Merge pull request #882 from PaddlePaddle/decoder

refactor raw ctc decoder into ctcdecoder, new join ctc/att decoder

deepspeech/decoders/README.MD

@@ -1,3 +0,0 @@
-# Reference
-* [Sequence Modeling With CTC](https://distill.pub/2017/ctc/)
-* [First-Pass Large Vocabulary Continuous Speech Recognition using Bi-Directional Recurrent DNNs](https://arxiv.org/pdf/1408.2873.pdf)

deepspeech/decoders/README.md

@@ -0,0 +1,13 @@
+# Decoders
+
+## Reference
+### CTC Prefix Beam Search
+* [Sequence Modeling With CTC](https://distill.pub/2017/ctc/)
+* [First-Pass Large Vocabulary Continuous Speech Recognition using Bi-Directional Recurrent DNNs](https://arxiv.org/pdf/1408.2873.pdf)
+
+### CTC Prefix Score & Join CTC/ATT One-passing Decoding
+* [Hybrid CTC/Attention Architecture for End-to-End Speech Recognition](http://www.ifp.illinois.edu/speech/speech_web_lg/slides/2019/watanabe_hybridCTCAttention_2017.pdf)
+* [Vectorized Beam Search for CTC-Attention-based Speech Recognition](https://www.isca-speech.org/archive/pdfs/interspeech_2019/seki19b_interspeech.pdf)
+
+### Streaming Join CTC/ATT Beam Search
+* [STREAMING TRANSFORMER ASR WITH BLOCKWISE SYNCHRONOUS BEAM SEARCH](https://arxiv.org/abs/2006.14941)

deepspeech/decoders/__init__.py

@@ -1,13 +1 @@
-# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+from .ctcdecoder import swig_wrapper 
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.

deepspeech/decoders/ctcdecoder/swig/__init__.py

@@ -0,0 +1,13 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

deepspeech/decoders/ctcdecoder/swig/ctc_beam_search_decoder.cpp

@@ -0,0 +1,243 @@
+// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ctc_beam_search_decoder.h"
+
+#include <algorithm>
+#include <cmath>
+#include <iostream>
+#include <limits>
+#include <map>
+#include <utility>
+
+#include "ThreadPool.h"
+#include "fst/fstlib.h"
+
+#include "decoder_utils.h"
+#include "path_trie.h"
+
+using FSTMATCH = fst::SortedMatcher<fst::StdVectorFst>;
+
+std::vector<std::pair<double, std::string>> ctc_beam_search_decoder(
+    const std::vector<std::vector<double>> &probs_seq,
+    const std::vector<std::string> &vocabulary,
+    size_t beam_size,
+    double cutoff_prob,
+    size_t cutoff_top_n,
+    Scorer *ext_scorer,
+    size_t blank_id) {
+    // dimension check
+    size_t num_time_steps = probs_seq.size();
+    for (size_t i = 0; i < num_time_steps; ++i) {
+        VALID_CHECK_EQ(probs_seq[i].size(),
+                       // vocabulary.size() + 1,
+                       vocabulary.size(),
+                       "The shape of probs_seq does not match with "
+                       "the shape of the vocabulary");
+    }
+    // assign space id
+    auto it = std::find(vocabulary.begin(), vocabulary.end(), kSPACE);
+    int space_id = it - vocabulary.begin();
+    // if no space in vocabulary
+    if ((size_t)space_id >= vocabulary.size()) {
+        space_id = -2;
+    }
+    // init prefixes' root
+    PathTrie root;
+    root.score = root.log_prob_b_prev = 0.0;
+    std::vector<PathTrie *> prefixes;
+    prefixes.push_back(&root);
+
+    if (ext_scorer != nullptr && !ext_scorer->is_character_based()) {
+        auto fst_dict =
+            static_cast<fst::StdVectorFst *>(ext_scorer->dictionary);
+        fst::StdVectorFst *dict_ptr = fst_dict->Copy(true);
+        root.set_dictionary(dict_ptr);
+        auto matcher = std::make_shared<FSTMATCH>(*dict_ptr, fst::MATCH_INPUT);
+        root.set_matcher(matcher);
+    }
+
+    // prefix search over time
+    for (size_t time_step = 0; time_step < num_time_steps; ++time_step) {
+        auto &prob = probs_seq[time_step];
+
+        float min_cutoff = -NUM_FLT_INF;
+        bool full_beam = false;
+        if (ext_scorer != nullptr) {
+            size_t num_prefixes = std::min(prefixes.size(), beam_size);
+            std::sort(prefixes.begin(),
+                      prefixes.begin() + num_prefixes,
+                      prefix_compare);
+            min_cutoff = prefixes[num_prefixes - 1]->score +
+                         std::log(prob[blank_id]) -
+                         std::max(0.0, ext_scorer->beta);
+            full_beam = (num_prefixes == beam_size);
+        }
+
+        std::vector<std::pair<size_t, float>> log_prob_idx =
+            get_pruned_log_probs(prob, cutoff_prob, cutoff_top_n);
+        // loop over chars
+        for (size_t index = 0; index < log_prob_idx.size(); index++) {
+            auto c = log_prob_idx[index].first;
+            auto log_prob_c = log_prob_idx[index].second;
+
+            for (size_t i = 0; i < prefixes.size() && i < beam_size; ++i) {
+                auto prefix = prefixes[i];
+                if (full_beam && log_prob_c + prefix->score < min_cutoff) {
+                    break;
+                }
+                // blank
+                if (c == blank_id) {
+                    prefix->log_prob_b_cur = log_sum_exp(
+                        prefix->log_prob_b_cur, log_prob_c + prefix->score);
+                    continue;
+                }
+                // repeated character
+                if (c == prefix->character) {
+                    prefix->log_prob_nb_cur =
+                        log_sum_exp(prefix->log_prob_nb_cur,
+                                    log_prob_c + prefix->log_prob_nb_prev);
+                }
+                // get new prefix
+                auto prefix_new = prefix->get_path_trie(c);
+
+                if (prefix_new != nullptr) {
+                    float log_p = -NUM_FLT_INF;
+
+                    if (c == prefix->character &&
+                        prefix->log_prob_b_prev > -NUM_FLT_INF) {
+                        log_p = log_prob_c + prefix->log_prob_b_prev;
+                    } else if (c != prefix->character) {
+                        log_p = log_prob_c + prefix->score;
+                    }
+
+                    // language model scoring
+                    if (ext_scorer != nullptr &&
+                        (c == space_id || ext_scorer->is_character_based())) {
+                        PathTrie *prefix_to_score = nullptr;
+                        // skip scoring the space
+                        if (ext_scorer->is_character_based()) {
+                            prefix_to_score = prefix_new;
+                        } else {
+                            prefix_to_score = prefix;
+                        }
+
+                        float score = 0.0;
+                        std::vector<std::string> ngram;
+                        ngram = ext_scorer->make_ngram(prefix_to_score);
+                        score = ext_scorer->get_log_cond_prob(ngram) *
+                                ext_scorer->alpha;
+                        log_p += score;
+                        log_p += ext_scorer->beta;
+                    }
+                    prefix_new->log_prob_nb_cur =
+                        log_sum_exp(prefix_new->log_prob_nb_cur, log_p);
+                }
+            }  // end of loop over prefix
+        }      // end of loop over vocabulary
+
+
+        prefixes.clear();
+        // update log probs
+        root.iterate_to_vec(prefixes);
+
+        // only preserve top beam_size prefixes
+        if (prefixes.size() >= beam_size) {
+            std::nth_element(prefixes.begin(),
+                             prefixes.begin() + beam_size,
+                             prefixes.end(),
+                             prefix_compare);
+            for (size_t i = beam_size; i < prefixes.size(); ++i) {
+                prefixes[i]->remove();
+            }
+        }
+    }  // end of loop over time
+
+    // score the last word of each prefix that doesn't end with space
+    if (ext_scorer != nullptr && !ext_scorer->is_character_based()) {
+        for (size_t i = 0; i < beam_size && i < prefixes.size(); ++i) {
+            auto prefix = prefixes[i];
+            if (!prefix->is_empty() && prefix->character != space_id) {
+                float score = 0.0;
+                std::vector<std::string> ngram = ext_scorer->make_ngram(prefix);
+                score =
+                    ext_scorer->get_log_cond_prob(ngram) * ext_scorer->alpha;
+                score += ext_scorer->beta;
+                prefix->score += score;
+            }
+        }
+    }
+
+    size_t num_prefixes = std::min(prefixes.size(), beam_size);
+    std::sort(
+        prefixes.begin(), prefixes.begin() + num_prefixes, prefix_compare);
+
+    // compute aproximate ctc score as the return score, without affecting the
+    // return order of decoding result. To delete when decoder gets stable.
+    for (size_t i = 0; i < beam_size && i < prefixes.size(); ++i) {
+        double approx_ctc = prefixes[i]->score;
+        if (ext_scorer != nullptr) {
+            std::vector<int> output;
+            prefixes[i]->get_path_vec(output);
+            auto prefix_length = output.size();
+            auto words = ext_scorer->split_labels(output);
+            // remove word insert
+            approx_ctc = approx_ctc - prefix_length * ext_scorer->beta;
+            // remove language model weight:
+            approx_ctc -=
+                (ext_scorer->get_sent_log_prob(words)) * ext_scorer->alpha;
+        }
+        prefixes[i]->approx_ctc = approx_ctc;
+    }
+
+    return get_beam_search_result(prefixes, vocabulary, beam_size);
+}
+
+
+std::vector<std::vector<std::pair<double, std::string>>>
+ctc_beam_search_decoder_batch(
+    const std::vector<std::vector<std::vector<double>>> &probs_split,
+    const std::vector<std::string> &vocabulary,
+    size_t beam_size,
+    size_t num_processes,
+    double cutoff_prob,
+    size_t cutoff_top_n,
+    Scorer *ext_scorer,
+    size_t blank_id) {
+    VALID_CHECK_GT(num_processes, 0, "num_processes must be nonnegative!");
+    // thread pool
+    ThreadPool pool(num_processes);
+    // number of samples
+    size_t batch_size = probs_split.size();
+
+    // enqueue the tasks of decoding
+    std::vector<std::future<std::vector<std::pair<double, std::string>>>> res;
+    for (size_t i = 0; i < batch_size; ++i) {
+        res.emplace_back(pool.enqueue(ctc_beam_search_decoder,
+                                      probs_split[i],
+                                      vocabulary,
+                                      beam_size,
+                                      cutoff_prob,
+                                      cutoff_top_n,
+                                      ext_scorer,
+                                      blank_id));
+    }
+
+    // get decoding results
+    std::vector<std::vector<std::pair<double, std::string>>> batch_results;
+    for (size_t i = 0; i < batch_size; ++i) {
+        batch_results.emplace_back(res[i].get());
+    }
+    return batch_results;
+}

deepspeech/decoders/ctcdecoder/swig/decoder_utils.h

@@ -0,0 +1,110 @@
+// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef DECODER_UTILS_H_
+#define DECODER_UTILS_H_
+
+#include <string>
+#include <utility>
+#include "fst/log.h"
+#include "path_trie.h"
+
+const std::string kSPACE = "<space>";
+const float NUM_FLT_INF = std::numeric_limits<float>::max();
+const float NUM_FLT_MIN = std::numeric_limits<float>::min();
+
+// inline function for validation check
+inline void check(
+    bool x, const char *expr, const char *file, int line, const char *err) {
+    if (!x) {
+        std::cout << "[" << file << ":" << line << "] ";
+        LOG(FATAL) << "\"" << expr << "\" check failed. " << err;
+    }
+}
+
+#define VALID_CHECK(x, info) \
+    check(static_cast<bool>(x), #x, __FILE__, __LINE__, info)
+#define VALID_CHECK_EQ(x, y, info) VALID_CHECK((x) == (y), info)
+#define VALID_CHECK_GT(x, y, info) VALID_CHECK((x) > (y), info)
+#define VALID_CHECK_LT(x, y, info) VALID_CHECK((x) < (y), info)
+
+
+// Function template for comparing two pairs
+template <typename T1, typename T2>
+bool pair_comp_first_rev(const std::pair<T1, T2> &a,
+                         const std::pair<T1, T2> &b) {
+    return a.first > b.first;
+}
+
+// Function template for comparing two pairs
+template <typename T1, typename T2>
+bool pair_comp_second_rev(const std::pair<T1, T2> &a,
+                          const std::pair<T1, T2> &b) {
+    return a.second > b.second;
+}
+
+// Return the sum of two probabilities in log scale
+template <typename T>
+T log_sum_exp(const T &x, const T &y) {
+    static T num_min = -std::numeric_limits<T>::max();
+    if (x <= num_min) return y;
+    if (y <= num_min) return x;
+    T xmax = std::max(x, y);
+    return std::log(std::exp(x - xmax) + std::exp(y - xmax)) + xmax;
+}
+
+// Get pruned probability vector for each time step's beam search
+std::vector<std::pair<size_t, float>> get_pruned_log_probs(
+    const std::vector<double> &prob_step,
+    double cutoff_prob,
+    size_t cutoff_top_n);
+
+// Get beam search result from prefixes in trie tree
+std::vector<std::pair<double, std::string>> get_beam_search_result(
+    const std::vector<PathTrie *> &prefixes,
+    const std::vector<std::string> &vocabulary,
+    size_t beam_size);
+
+// Functor for prefix comparsion
+bool prefix_compare(const PathTrie *x, const PathTrie *y);
+
+/* Get length of utf8 encoding string
+ * See: http://stackoverflow.com/a/4063229
+ */
+size_t get_utf8_str_len(const std::string &str);
+
+/* Split a string into a list of strings on a given string
+ * delimiter. NB: delimiters on beginning / end of string are
+ * trimmed. Eg, "FooBarFoo" split on "Foo" returns ["Bar"].
+ */
+std::vector<std::string> split_str(const std::string &s,
+                                   const std::string &delim);
+
+/* Splits string into vector of strings representing
+ * UTF-8 characters (not same as chars)
+ */
+std::vector<std::string> split_utf8_str(const std::string &str);
+
+// Add a word in index to the dicionary of fst
+void add_word_to_fst(const std::vector<int> &word,
+                     fst::StdVectorFst *dictionary);
+
+// Add a word in string to dictionary
+bool add_word_to_dictionary(
+    const std::string &word,
+    const std::unordered_map<std::string, int> &char_map,
+    bool add_space,
+    int SPACE_ID,
+    fst::StdVectorFst *dictionary);
+#endif  // DECODER_UTILS_H

deepspeech/decoders/ctcdecoder/swig/scorer.cpp

@@ -0,0 +1,244 @@
+// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "scorer.h"
+
+#include <unistd.h>
+#include <iostream>
+
+#include "lm/config.hh"
+#include "lm/model.hh"
+#include "lm/state.hh"
+#include "util/string_piece.hh"
+#include "util/tokenize_piece.hh"
+
+#include "decoder_utils.h"
+
+using namespace lm::ngram;
+
+Scorer::Scorer(double alpha,
+               double beta,
+               const std::string& lm_path,
+               const std::vector<std::string>& vocab_list) {
+    this->alpha = alpha;
+    this->beta = beta;
+
+    dictionary = nullptr;
+    is_character_based_ = true;
+    language_model_ = nullptr;
+
+    max_order_ = 0;
+    dict_size_ = 0;
+    SPACE_ID_ = -1;
+
+    setup(lm_path, vocab_list);
+}
+
+Scorer::~Scorer() {
+    if (language_model_ != nullptr) {
+        delete static_cast<lm::base::Model*>(language_model_);
+    }
+    if (dictionary != nullptr) {
+        delete static_cast<fst::StdVectorFst*>(dictionary);
+    }
+}
+
+void Scorer::setup(const std::string& lm_path,
+                   const std::vector<std::string>& vocab_list) {
+    // load language model
+    load_lm(lm_path);
+    // set char map for scorer
+    set_char_map(vocab_list);
+    // fill the dictionary for FST
+    if (!is_character_based()) {
+        fill_dictionary(true);
+    }
+}
+
+void Scorer::load_lm(const std::string& lm_path) {
+    const char* filename = lm_path.c_str();
+    VALID_CHECK_EQ(access(filename, F_OK), 0, "Invalid language model path");
+
+    RetriveStrEnumerateVocab enumerate;
+    lm::ngram::Config config;
+    config.enumerate_vocab = &enumerate;
+    language_model_ = lm::ngram::LoadVirtual(filename, config);
+    max_order_ = static_cast<lm::base::Model*>(language_model_)->Order();
+    vocabulary_ = enumerate.vocabulary;
+    for (size_t i = 0; i < vocabulary_.size(); ++i) {
+        if (is_character_based_ && vocabulary_[i] != UNK_TOKEN &&
+            vocabulary_[i] != START_TOKEN && vocabulary_[i] != END_TOKEN &&
+            get_utf8_str_len(enumerate.vocabulary[i]) > 1) {
+            is_character_based_ = false;
+        }
+    }
+}
+
+double Scorer::get_log_cond_prob(const std::vector<std::string>& words) {
+    lm::base::Model* model = static_cast<lm::base::Model*>(language_model_);
+    double cond_prob;
+    lm::ngram::State state, tmp_state, out_state;
+    // avoid to inserting <s> in begin
+    model->NullContextWrite(&state);
+    for (size_t i = 0; i < words.size(); ++i) {
+        lm::WordIndex word_index = model->BaseVocabulary().Index(words[i]);
+        // encounter OOV
+        if (word_index == 0) {
+            return OOV_SCORE;
+        }
+        cond_prob = model->BaseScore(&state, word_index, &out_state);
+        tmp_state = state;
+        state = out_state;
+        out_state = tmp_state;
+    }
+    // return  log10 prob
+    return cond_prob;
+}
+
+double Scorer::get_sent_log_prob(const std::vector<std::string>& words) {
+    std::vector<std::string> sentence;
+    if (words.size() == 0) {
+        for (size_t i = 0; i < max_order_; ++i) {
+            sentence.push_back(START_TOKEN);
+        }
+    } else {
+        for (size_t i = 0; i < max_order_ - 1; ++i) {
+            sentence.push_back(START_TOKEN);
+        }
+        sentence.insert(sentence.end(), words.begin(), words.end());
+    }
+    sentence.push_back(END_TOKEN);
+    return get_log_prob(sentence);
+}
+
+double Scorer::get_log_prob(const std::vector<std::string>& words) {
+    assert(words.size() > max_order_);
+    double score = 0.0;
+    for (size_t i = 0; i < words.size() - max_order_ + 1; ++i) {
+        std::vector<std::string> ngram(words.begin() + i,
+                                       words.begin() + i + max_order_);
+        score += get_log_cond_prob(ngram);
+    }
+    return score;
+}
+
+void Scorer::reset_params(float alpha, float beta) {
+    this->alpha = alpha;
+    this->beta = beta;
+}
+
+std::string Scorer::vec2str(const std::vector<int>& input) {
+    std::string word;
+    for (auto ind : input) {
+        word += char_list_[ind];
+    }
+    return word;
+}
+
+std::vector<std::string> Scorer::split_labels(const std::vector<int>& labels) {
+    if (labels.empty()) return {};
+
+    std::string s = vec2str(labels);
+    std::vector<std::string> words;
+    if (is_character_based_) {
+        words = split_utf8_str(s);
+    } else {
+        words = split_str(s, " ");
+    }
+    return words;
+}
+
+void Scorer::set_char_map(const std::vector<std::string>& char_list) {
+    char_list_ = char_list;
+    char_map_.clear();
+
+    // Set the char map for the FST for spelling correction
+    for (size_t i = 0; i < char_list_.size(); i++) {
+        if (char_list_[i] == kSPACE) {
+            SPACE_ID_ = i;
+        }
+        // The initial state of FST is state 0, hence the index of chars in
+        // the FST should start from 1 to avoid the conflict with the initial
+        // state, otherwise wrong decoding results would be given.
+        char_map_[char_list_[i]] = i + 1;
+    }
+}
+
+std::vector<std::string> Scorer::make_ngram(PathTrie* prefix) {
+    std::vector<std::string> ngram;
+    PathTrie* current_node = prefix;
+    PathTrie* new_node = nullptr;
+
+    for (int order = 0; order < max_order_; order++) {
+        std::vector<int> prefix_vec;
+
+        if (is_character_based_) {
+            new_node = current_node->get_path_vec(prefix_vec, SPACE_ID_, 1);
+            current_node = new_node;
+        } else {
+            new_node = current_node->get_path_vec(prefix_vec, SPACE_ID_);
+            current_node = new_node->parent;  // Skipping spaces
+        }
+
+        // reconstruct word
+        std::string word = vec2str(prefix_vec);
+        ngram.push_back(word);
+
+        if (new_node->character == -1) {
+            // No more spaces, but still need order
+            for (int i = 0; i < max_order_ - order - 1; i++) {
+                ngram.push_back(START_TOKEN);
+            }
+            break;
+        }
+    }
+    std::reverse(ngram.begin(), ngram.end());
+    return ngram;
+}
+
+void Scorer::fill_dictionary(bool add_space) {
+    fst::StdVectorFst dictionary;
+    // For each unigram convert to ints and put in trie
+    int dict_size = 0;
+    for (const auto& word : vocabulary_) {
+        bool added = add_word_to_dictionary(
+            word, char_map_, add_space, SPACE_ID_ + 1, &dictionary);
+        dict_size += added ? 1 : 0;
+    }
+
+    dict_size_ = dict_size;
+
+    /* Simplify FST
+
+     * This gets rid of "epsilon" transitions in the FST.
+     * These are transitions that don't require a string input to be taken.
+     * Getting rid of them is necessary to make the FST determinisitc, but
+     * can greatly increase the size of the FST
+     */
+    fst::RmEpsilon(&dictionary);
+    fst::StdVectorFst* new_dict = new fst::StdVectorFst;
+
+    /* This makes the FST deterministic, meaning for any string input there's
+     * only one possible state the FST could be in.  It is assumed our
+     * dictionary is deterministic when using it.
+     * (lest we'd have to check for multiple transitions at each state)
+     */
+    fst::Determinize(dictionary, new_dict);
+
+    /* Finds the simplest equivalent fst. This is unnecessary but decreases
+     * memory usage of the dictionary
+     */
+    fst::Minimize(new_dict);
+    this->dictionary = new_dict;
+}

deepspeech/decoders/ctcdecoder/swig/setup.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+
+if [ ! -d kenlm ]; then
+    git clone https://github.com/kpu/kenlm.git
+    cd kenlm/
+    git checkout df2d717e95183f79a90b2fa6e4307083a351ca6a
+    cd ..
+    echo -e "\n"
+fi
+
+if [ ! -d openfst-1.6.3 ]; then
+    echo "Download and extract openfst ..."
+    wget http://www.openfst.org/twiki/pub/FST/FstDownload/openfst-1.6.3.tar.gz
+    tar -xzvf openfst-1.6.3.tar.gz
+    echo -e "\n"
+fi
+
+if [ ! -d ThreadPool ]; then
+    git clone https://github.com/progschj/ThreadPool.git
+    echo -e "\n"
+fi
+
+echo "Install decoders ..."
+python3 setup.py install --num_processes 4

deepspeech/decoders/scores/ctc.py

@@ -0,0 +1,158 @@
+"""ScorerInterface implementation for CTC."""
+
+import numpy as np
+import paddle
+
+from .ctc_prefix_score import CTCPrefixScore
+from .ctc_prefix_score import CTCPrefixScorePD
+from .scorer_interface import BatchPartialScorerInterface
+
+
+class CTCPrefixScorer(BatchPartialScorerInterface):
+    """Decoder interface wrapper for CTCPrefixScore."""
+
+    def __init__(self, ctc: paddle.nn.Layer, eos: int):
+        """Initialize class.
+
+        Args:
+            ctc (paddle.nn.Layer): The CTC implementation.
+                For example, :class:`deepspeech.modules.ctc.CTC`
+            eos (int): The end-of-sequence id.
+
+        """
+        self.ctc = ctc
+        self.eos = eos
+        self.impl = None
+
+    def init_state(self, x: paddle.Tensor):
+        """Get an initial state for decoding.
+
+        Args:
+            x (paddle.Tensor): The encoded feature tensor
+
+        Returns: initial state
+
+        """
+        logp = self.ctc.log_softmax(x.unsqueeze(0)).squeeze(0).numpy()
+        # TODO(karita): use CTCPrefixScorePD
+        self.impl = CTCPrefixScore(logp, 0, self.eos, np)
+        return 0, self.impl.initial_state()
+
+    def select_state(self, state, i, new_id=None):
+        """Select state with relative ids in the main beam search.
+
+        Args:
+            state: Decoder state for prefix tokens
+            i (int): Index to select a state in the main beam search
+            new_id (int): New label id to select a state if necessary
+
+        Returns:
+            state: pruned state
+
+        """
+        if type(state) == tuple:
+            if len(state) == 2:  # for CTCPrefixScore
+                sc, st = state
+                return sc[i], st[i]
+            else:  # for CTCPrefixScorePD (need new_id > 0)
+                r, log_psi, f_min, f_max, scoring_idmap = state
+                s = log_psi[i, new_id].expand(log_psi.size(1))
+                if scoring_idmap is not None:
+                    return r[:, :, i, scoring_idmap[i, new_id]], s, f_min, f_max
+                else:
+                    return r[:, :, i, new_id], s, f_min, f_max
+        return None if state is None else state[i]
+
+    def score_partial(self, y, ids, state, x):
+        """Score new token.
+
+        Args:
+            y (paddle.Tensor): 1D prefix token
+            next_tokens (paddle.Tensor): paddle.int64 next token to score
+            state: decoder state for prefix tokens
+            x (paddle.Tensor): 2D encoder feature that generates ys
+
+        Returns:
+            tuple[paddle.Tensor, Any]:
+                Tuple of a score tensor for y that has a shape `(len(next_tokens),)`
+                and next state for ys
+
+        """
+        prev_score, state = state
+        presub_score, new_st = self.impl(y.cpu(), ids.cpu(), state)
+        tscore = paddle.to_tensor(
+            presub_score - prev_score, place=x.place, dtype=x.dtype
+        )
+        return tscore, (presub_score, new_st)
+
+    def batch_init_state(self, x: paddle.Tensor):
+        """Get an initial state for decoding.
+
+        Args:
+            x (paddle.Tensor): The encoded feature tensor
+
+        Returns: initial state
+
+        """
+        logp = self.ctc.log_softmax(x.unsqueeze(0))  # assuming batch_size = 1
+        xlen = paddle.to_tensor([logp.size(1)])
+        self.impl = CTCPrefixScorePD(logp, xlen, 0, self.eos)
+        return None
+
+    def batch_score_partial(self, y, ids, state, x):
+        """Score new token.
+
+        Args:
+            y (paddle.Tensor): 1D prefix token
+            ids (paddle.Tensor): paddle.int64 next token to score
+            state: decoder state for prefix tokens
+            x (paddle.Tensor): 2D encoder feature that generates ys
+
+        Returns:
+            tuple[paddle.Tensor, Any]:
+                Tuple of a score tensor for y that has a shape `(len(next_tokens),)`
+                and next state for ys
+
+        """
+        batch_state = (
+            (
+                paddle.stack([s[0] for s in state], axis=2),
+                paddle.stack([s[1] for s in state]),
+                state[0][2],
+                state[0][3],
+            )
+            if state[0] is not None
+            else None
+        )
+        return self.impl(y, batch_state, ids)
+
+    def extend_prob(self, x: paddle.Tensor):
+        """Extend probs for decoding.
+
+        This extension is for streaming decoding
+        as in Eq (14) in https://arxiv.org/abs/2006.14941
+
+        Args:
+            x (paddle.Tensor): The encoded feature tensor
+
+        """
+        logp = self.ctc.log_softmax(x.unsqueeze(0))
+        self.impl.extend_prob(logp)
+
+    def extend_state(self, state):
+        """Extend state for decoding.
+
+        This extension is for streaming decoding
+        as in Eq (14) in https://arxiv.org/abs/2006.14941
+
+        Args:
+            state: The states of hyps
+
+        Returns: exteded state
+
+        """
+        new_state = []
+        for s in state:
+            new_state.append(self.impl.extend_state(s))
+
+        return new_state

deepspeech/decoders/scores/ctc_prefix_score.py

@@ -0,0 +1,356 @@
+#!/usr/bin/env python3
+
+# Copyright 2018 Mitsubishi Electric Research Labs (Takaaki Hori)
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+import paddle
+
+import numpy as np
+import six
+
+
+class CTCPrefixScorePD():
+    """Batch processing of CTCPrefixScore
+
+    which is based on Algorithm 2 in WATANABE et al.
+    "HYBRID CTC/ATTENTION ARCHITECTURE FOR END-TO-END SPEECH RECOGNITION,"
+    but extended to efficiently compute the label probablities for multiple
+    hypotheses simultaneously
+    See also Seki et al. "Vectorized Beam Search for CTC-Attention-Based
+    Speech Recognition," In INTERSPEECH (pp. 3825-3829), 2019.
+    """
+
+    def __init__(self, x, xlens, blank, eos, margin=0):
+        """Construct CTC prefix scorer
+
+        `margin` is M in eq.(22,23)
+
+        :param paddle.Tensor x: input label posterior sequences (B, T, O)
+        :param paddle.Tensor xlens: input lengths (B,)
+        :param int blank: blank label id
+        :param int eos: end-of-sequence id
+        :param int margin: margin parameter for windowing (0 means no windowing)
+        """
+        # In the comment lines,
+        # we assume T: input_length, B: batch size, W: beam width, O: output dim.
+        self.logzero = -10000000000.0
+        self.blank = blank
+        self.eos = eos
+        self.batch = x.size(0)
+        self.input_length = x.size(1)
+        self.odim = x.size(2)
+        self.dtype = x.dtype
+
+        # Pad the rest of posteriors in the batch
+        # TODO(takaaki-hori): need a better way without for-loops
+        for i, l in enumerate(xlens):
+            if l < self.input_length:
+                x[i, l:, :] = self.logzero
+                x[i, l:, blank] = 0
+        # Reshape input x
+        xn = x.transpose([1, 0, 2])  # (B, T, O) -> (T, B, O)
+        xb = xn[:, :, self.blank].unsqueeze(2).expand(-1, -1, self.odim) # (T,B,O)
+        self.x = paddle.stack([xn, xb])  # (2, T, B, O)
+        self.end_frames = paddle.to_tensor(xlens) - 1 # (B,)
+
+        # Setup CTC windowing
+        self.margin = margin
+        if margin > 0:
+            self.frame_ids = paddle.arange(self.input_length, dtype=self.dtype)
+        # Base indices for index conversion
+        # B idx, hyp idx. shape (B*W, 1)
+        self.idx_bh = None 
+        # B idx. shape (B,)
+        self.idx_b = paddle.arange(self.batch)
+        # B idx, O idx. shape (B, 1)
+        self.idx_bo = (self.idx_b * self.odim).unsqueeze(1)
+
+    def __call__(self, y, state, scoring_ids=None, att_w=None):
+        """Compute CTC prefix scores for next labels
+
+        :param list y: prefix label sequences
+        :param tuple state: previous CTC state
+        :param paddle.Tensor scoring_ids: selected next ids to score (BW, O'), O' <= O
+        :param paddle.Tensor att_w: attention weights to decide CTC window
+        :return new_state, ctc_local_scores (BW, O)
+        """
+        output_length = len(y[0]) - 1  # ignore sos
+        last_ids = [yi[-1] for yi in y]  # last output label ids
+        n_bh = len(last_ids)  # batch * hyps
+        n_hyps = n_bh // self.batch  # assuming each utterance has the same # of hyps
+        self.scoring_num = scoring_ids.size(-1) if scoring_ids is not None else 0
+        # prepare state info
+        if state is None:
+            r_prev = paddle.full(
+                (self.input_length, 2, self.batch, n_hyps),
+                self.logzero,
+                dtype=self.dtype,
+            ) # (T, 2, B, W)
+            r_prev[:, 1] = paddle.cumsum(self.x[0, :, :, self.blank], 0).unsqueeze(2)
+            r_prev = r_prev.view(-1, 2, n_bh) # (T, 2, BW)
+            s_prev = 0.0  # score
+            f_min_prev = 0 # eq. 22-23
+            f_max_prev = 1 # eq. 22-23
+        else:
+            r_prev, s_prev, f_min_prev, f_max_prev = state
+
+        # select input dimensions for scoring
+        if self.scoring_num > 0:
+            # (BW, O)
+            scoring_idmap = paddle.full((n_bh, self.odim), -1, dtype=paddle.long)
+            snum = self.scoring_num
+            if self.idx_bh is None or n_bh > len(self.idx_bh):
+                self.idx_bh = paddle.arange(n_bh).view(-1, 1) # (BW, 1)
+            scoring_idmap[self.idx_bh[:n_bh], scoring_ids] = paddle.arange(snum)
+            scoring_idx = (
+                scoring_ids + self.idx_bo.repeat(1, n_hyps).view(-1, 1) # (BW,1)
+            ).view(-1) # (BWO)
+            # x_ shape (2, T, B*W, O)
+            x_ = paddle.index_select(
+                self.x.view(2, -1, self.batch * self.odim), scoring_idx, 2
+            ).view(2, -1, n_bh, snum)
+        else:
+            scoring_ids = None
+            scoring_idmap = None
+            snum = self.odim
+            # x_ shape (2, T, B*W, O)
+            x_ = self.x.unsqueeze(3).repeat(1, 1, 1, n_hyps, 1).view(2, -1, n_bh, snum)
+
+        # new CTC forward probs are prepared as a (T x 2 x BW x S) tensor
+        # that corresponds to r_t^n(h) and r_t^b(h) in a batch.
+        r = paddle.full(
+            (self.input_length, 2, n_bh, snum),
+            self.logzero,
+            dtype=self.dtype,
+        )
+        if output_length == 0:
+            r[0, 0] = x_[0, 0]
+
+        r_sum = paddle.logsumexp(r_prev, 1) #(T,BW)
+        log_phi = r_sum.unsqueeze(2).repeat(1, 1, snum) # (T, BW, O)
+        if scoring_ids is not None:
+            for idx in range(n_bh):
+                pos = scoring_idmap[idx, last_ids[idx]]
+                if pos >= 0:
+                    log_phi[:, idx, pos] = r_prev[:, 1, idx]
+        else:
+            for idx in range(n_bh):
+                log_phi[:, idx, last_ids[idx]] = r_prev[:, 1, idx]
+
+        # decide start and end frames based on attention weights
+        if att_w is not None and self.margin > 0:
+            f_arg = paddle.matmul(att_w, self.frame_ids)
+            f_min = max(int(f_arg.min().cpu()), f_min_prev)
+            f_max = max(int(f_arg.max().cpu()), f_max_prev)
+            start = min(f_max_prev, max(f_min - self.margin, output_length, 1))
+            end = min(f_max + self.margin, self.input_length)
+        else:
+            f_min = f_max = 0
+            # if one frame one out, the output_length is the eating frame num now.
+            start = max(output_length, 1)
+            end = self.input_length
+
+        # compute forward probabilities log(r_t^n(h)) and log(r_t^b(h))
+        for t in range(start, end):
+            rp = r[t - 1] # (2 x BW x O') 
+            rr = paddle.stack([rp[0], log_phi[t - 1], rp[0], rp[1]]).view(
+                2, 2, n_bh, snum
+            ) # (2,2,BW,O')
+            r[t] = paddle.logsumexp(rr, 1) + x_[:, t]
+
+        # compute log prefix probabilities log(psi)
+        log_phi_x = paddle.concat((log_phi[0].unsqueeze(0), log_phi[:-1]), axis=0) + x_[0]
+        if scoring_ids is not None:
+            log_psi = paddle.full((n_bh, self.odim), self.logzero, dtype=self.dtype)
+            log_psi_ = paddle.logsumexp(
+                paddle.concat((log_phi_x[start:end], r[start - 1, 0].unsqueeze(0)), axis=0),
+                axis=0,
+            )
+            for si in range(n_bh):
+                log_psi[si, scoring_ids[si]] = log_psi_[si]
+        else:
+            log_psi = paddle.logsumexp(
+                paddle.concat((log_phi_x[start:end], r[start - 1, 0].unsqueeze(0)), axis=0),
+                axis=0,
+            )
+
+        for si in range(n_bh):
+            log_psi[si, self.eos] = r_sum[self.end_frames[si // n_hyps], si]
+
+        # exclude blank probs
+        log_psi[:, self.blank] = self.logzero
+
+        return (log_psi - s_prev), (r, log_psi, f_min, f_max, scoring_idmap)
+
+    def index_select_state(self, state, best_ids):
+        """Select CTC states according to best ids
+
+        :param state    : CTC state
+        :param best_ids : index numbers selected by beam pruning (B, W)
+        :return selected_state
+        """
+        r, s, f_min, f_max, scoring_idmap = state
+        # convert ids to BHO space
+        n_bh = len(s)
+        n_hyps = n_bh // self.batch
+        vidx = (best_ids + (self.idx_b * (n_hyps * self.odim)).view(-1, 1)).view(-1)
+        # select hypothesis scores
+        s_new = paddle.index_select(s.view(-1), vidx, 0)
+        s_new = s_new.view(-1, 1).repeat(1, self.odim).view(n_bh, self.odim)
+        # convert ids to BHS space (S: scoring_num)
+        if scoring_idmap is not None:
+            snum = self.scoring_num
+            hyp_idx = (best_ids // self.odim + (self.idx_b * n_hyps).view(-1, 1)).view(
+                -1
+            )
+            label_ids = paddle.fmod(best_ids, self.odim).view(-1)
+            score_idx = scoring_idmap[hyp_idx, label_ids]
+            score_idx[score_idx == -1] = 0
+            vidx = score_idx + hyp_idx * snum
+        else:
+            snum = self.odim
+        # select forward probabilities
+        r_new = paddle.index_select(r.view(-1, 2, n_bh * snum), vidx, 2).view(
+            -1, 2, n_bh
+        )
+        return r_new, s_new, f_min, f_max
+
+    def extend_prob(self, x):
+        """Extend CTC prob.
+
+        :param paddle.Tensor x: input label posterior sequences (B, T, O)
+        """
+
+        if self.x.shape[1] < x.shape[1]:  # self.x (2,T,B,O); x (B,T,O)
+            # Pad the rest of posteriors in the batch
+            # TODO(takaaki-hori): need a better way without for-loops
+            xlens = [x.size(1)]
+            for i, l in enumerate(xlens):
+                if l < self.input_length:
+                    x[i, l:, :] = self.logzero
+                    x[i, l:, self.blank] = 0
+            tmp_x = self.x
+            xn = x.transpose([1, 0, 2])  # (B, T, O) -> (T, B, O)
+            xb = xn[:, :, self.blank].unsqueeze(2).expand(-1, -1, self.odim)
+            self.x = paddle.stack([xn, xb])  # (2, T, B, O)
+            self.x[:, : tmp_x.shape[1], :, :] = tmp_x
+            self.input_length = x.size(1)
+            self.end_frames = paddle.to_tensor(xlens) - 1
+
+    def extend_state(self, state):
+        """Compute CTC prefix state.
+
+
+        :param state    : CTC state
+        :return ctc_state
+        """
+
+        if state is None:
+            # nothing to do
+            return state
+        else:
+            r_prev, s_prev, f_min_prev, f_max_prev = state
+
+            r_prev_new = paddle.full(
+                (self.input_length, 2),
+                self.logzero,
+                dtype=self.dtype,
+            )
+            start = max(r_prev.shape[0], 1)
+            r_prev_new[0:start] = r_prev
+            for t in range(start, self.input_length):
+                r_prev_new[t, 1] = r_prev_new[t - 1, 1] + self.x[0, t, :, self.blank]
+
+            return (r_prev_new, s_prev, f_min_prev, f_max_prev)
+
+
+class CTCPrefixScore():
+    """Compute CTC label sequence scores
+
+    which is based on Algorithm 2 in WATANABE et al.
+    "HYBRID CTC/ATTENTION ARCHITECTURE FOR END-TO-END SPEECH RECOGNITION,"
+    but extended to efficiently compute the probablities of multiple labels
+    simultaneously
+    """
+
+    def __init__(self, x, blank, eos, xp):
+        self.xp = xp
+        self.logzero = -10000000000.0
+        self.blank = blank
+        self.eos = eos
+        self.input_length = len(x)
+        self.x = x # (T, O)
+
+    def initial_state(self):
+        """Obtain an initial CTC state
+
+        :return: CTC state
+        """
+        # initial CTC state is made of a frame x 2 tensor that corresponds to
+        # r_t^n(<sos>) and r_t^b(<sos>), where 0 and 1 of axis=1 represent
+        # superscripts n and b (non-blank and blank), respectively.
+        # r shape (T, 2)
+        r = self.xp.full((self.input_length, 2), self.logzero, dtype=np.float32)
+        r[0, 1] = self.x[0, self.blank]
+        for i in six.moves.range(1, self.input_length):
+            r[i, 1] = r[i - 1, 1] + self.x[i, self.blank]
+        return r
+
+    def __call__(self, y, cs, r_prev):
+        """Compute CTC prefix scores for next labels
+
+        :param y     : prefix label sequence
+        :param cs    : array of next labels
+        :param r_prev: previous CTC state
+        :return ctc_scores, ctc_states
+        """
+        # initialize CTC states
+        output_length = len(y) - 1  # ignore sos
+        # new CTC states are prepared as a frame x (n or b) x n_labels tensor
+        # that corresponds to r_t^n(h) and r_t^b(h).
+        # r shape (T, 2, n_labels)
+        r = self.xp.ndarray((self.input_length, 2, len(cs)), dtype=np.float32)
+        xs = self.x[:, cs]
+        if output_length == 0:
+            r[0, 0] = xs[0]
+            r[0, 1] = self.logzero
+        else:
+            r[output_length - 1] = self.logzero
+
+        # prepare forward probabilities for the last label
+        r_sum = self.xp.logaddexp(
+            r_prev[:, 0], r_prev[:, 1]
+        )  # log(r_t^n(g) + r_t^b(g))
+        last = y[-1]
+        if output_length > 0 and last in cs:
+            log_phi = self.xp.ndarray((self.input_length, len(cs)), dtype=np.float32)
+            for i in six.moves.range(len(cs)):
+                log_phi[:, i] = r_sum if cs[i] != last else r_prev[:, 1]
+        else:
+            log_phi = r_sum
+
+        # compute forward probabilities log(r_t^n(h)), log(r_t^b(h)),
+        # and log prefix probabilities log(psi)
+        start = max(output_length, 1)
+        log_psi = r[start - 1, 0]
+        for t in six.moves.range(start, self.input_length):
+            r[t, 0] = self.xp.logaddexp(r[t - 1, 0], log_phi[t - 1]) + xs[t]
+            r[t, 1] = (
+                self.xp.logaddexp(r[t - 1, 0], r[t - 1, 1]) + self.x[t, self.blank]
+            )
+            log_psi = self.xp.logaddexp(log_psi, log_phi[t - 1] + xs[t])
+
+        # get P(...eos|X) that ends with the prefix itself
+        eos_pos = self.xp.where(cs == self.eos)[0]
+        if len(eos_pos) > 0:
+            log_psi[eos_pos] = r_sum[-1]  # log(r_T^n(g) + r_T^b(g))
+
+        # exclude blank probs
+        blank_pos = self.xp.where(cs == self.blank)[0]
+        if len(blank_pos) > 0:
+            log_psi[blank_pos] = self.logzero
+
+        # return the log prefix probability and CTC states, where the label axis
+        # of the CTC states is moved to the first axis to slice it easily
+        # log_psi shape (n_labels,), state shape (n_labels, T, 2)
+        return log_psi, self.xp.rollaxis(r, 2)

deepspeech/decoders/scores/length_bonus.py

@@ -0,0 +1,61 @@
+"""Length bonus module."""
+from typing import Any
+from typing import List
+from typing import Tuple
+
+import paddle
+
+from .scorer_interface import BatchScorerInterface
+
+
+class LengthBonus(BatchScorerInterface):
+    """Length bonus in beam search."""
+
+    def __init__(self, n_vocab: int):
+        """Initialize class.
+
+        Args:
+            n_vocab (int): The number of tokens in vocabulary for beam search
+
+        """
+        self.n = n_vocab
+
+    def score(self, y, state, x):
+        """Score new token.
+
+        Args:
+            y (paddle.Tensor): 1D paddle.int64 prefix tokens.
+            state: Scorer state for prefix tokens
+            x (paddle.Tensor): 2D encoder feature that generates ys.
+
+        Returns:
+            tuple[paddle.Tensor, Any]: Tuple of
+                paddle.float32 scores for next token (n_vocab)
+                and None
+
+        """
+        return paddle.to_tensor([1.0], place=x.place, dtype=x.dtype).expand(self.n), None
+
+    def batch_score(
+        self, ys: paddle.Tensor, states: List[Any], xs: paddle.Tensor
+    ) -> Tuple[paddle.Tensor, List[Any]]:
+        """Score new token batch.
+
+        Args:
+            ys (paddle.Tensor): paddle.int64 prefix tokens (n_batch, ylen).
+            states (List[Any]): Scorer states for prefix tokens.
+            xs (paddle.Tensor):
+                The encoder feature that generates ys (n_batch, xlen, n_feat).
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        return (
+            paddle.to_tensor([1.0], place=xs.place, dtype=xs.dtype).expand(
+                ys.shape[0], self.n
+            ),
+            None,
+        )

deepspeech/decoders/scores/ngram.py

@@ -0,0 +1,102 @@
+"""Ngram lm implement."""
+
+from abc import ABC
+
+import kenlm
+import paddle
+
+from .scorer_interface import BatchScorerInterface
+from .scorer_interface import PartialScorerInterface
+
+
+class Ngrambase(ABC):
+    """Ngram base implemented through ScorerInterface."""
+
+    def __init__(self, ngram_model, token_list):
+        """Initialize Ngrambase.
+
+        Args:
+            ngram_model: ngram model path
+            token_list: token list from dict or model.json
+
+        """
+        self.chardict = [x if x != "<eos>" else "</s>" for x in token_list]
+        self.charlen = len(self.chardict)
+        self.lm = kenlm.LanguageModel(ngram_model)
+        self.tmpkenlmstate = kenlm.State()
+
+    def init_state(self, x):
+        """Initialize tmp state."""
+        state = kenlm.State()
+        self.lm.NullContextWrite(state)
+        return state
+
+    def score_partial_(self, y, next_token, state, x):
+        """Score interface for both full and partial scorer.
+
+        Args:
+            y: previous char
+            next_token: next token need to be score
+            state: previous state
+            x: encoded feature
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        out_state = kenlm.State()
+        ys = self.chardict[y[-1]] if y.shape[0] > 1 else "<s>"
+        self.lm.BaseScore(state, ys, out_state)
+        scores = paddle.empty_like(next_token, dtype=x.dtype)
+        for i, j in enumerate(next_token):
+            scores[i] = self.lm.BaseScore(
+                out_state, self.chardict[j], self.tmpkenlmstate
+            )
+        return scores, out_state
+
+
+class NgramFullScorer(Ngrambase, BatchScorerInterface):
+    """Fullscorer for ngram."""
+
+    def score(self, y, state, x):
+        """Score interface for both full and partial scorer.
+
+        Args:
+            y: previous char
+            state: previous state
+            x: encoded feature
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        return self.score_partial_(y, paddle.to_tensor(range(self.charlen)), state, x)
+
+
+class NgramPartScorer(Ngrambase, PartialScorerInterface):
+    """Partialscorer for ngram."""
+
+    def score_partial(self, y, next_token, state, x):
+        """Score interface for both full and partial scorer.
+
+        Args:
+            y: previous char
+            next_token: next token need to be score
+            state: previous state
+            x: encoded feature
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        return self.score_partial_(y, next_token, state, x)
+
+    def select_state(self, state, i):
+        """Empty select state for scorer interface."""
+        return state

deepspeech/decoders/scores/score_interface.py

@@ -0,0 +1,188 @@
+"""Scorer interface module."""
+
+from typing import Any
+from typing import List
+from typing import Tuple
+
+import paddle
+import warnings
+
+
+class ScorerInterface:
+    """Scorer interface for beam search.
+
+    The scorer performs scoring of the all tokens in vocabulary.
+
+    Examples:
+        * Search heuristics
+            * :class:`scorers.length_bonus.LengthBonus`
+        * Decoder networks of the sequence-to-sequence models
+            * :class:`transformer.decoder.Decoder`
+            * :class:`rnn.decoders.Decoder`
+        * Neural language models
+            * :class:`lm.transformer.TransformerLM`
+            * :class:`lm.default.DefaultRNNLM`
+            * :class:`lm.seq_rnn.SequentialRNNLM`
+
+    """
+
+    def init_state(self, x: paddle.Tensor) -> Any:
+        """Get an initial state for decoding (optional).
+
+        Args:
+            x (paddle.Tensor): The encoded feature tensor
+
+        Returns: initial state
+
+        """
+        return None
+
+    def select_state(self, state: Any, i: int, new_id: int = None) -> Any:
+        """Select state with relative ids in the main beam search.
+
+        Args:
+            state: Decoder state for prefix tokens
+            i (int): Index to select a state in the main beam search
+            new_id (int): New label index to select a state if necessary
+
+        Returns:
+            state: pruned state
+
+        """
+        return None if state is None else state[i]
+
+    def score(
+        self, y: paddle.Tensor, state: Any, x: paddle.Tensor
+    ) -> Tuple[paddle.Tensor, Any]:
+        """Score new token (required).
+
+        Args:
+            y (paddle.Tensor): 1D paddle.int64 prefix tokens.
+            state: Scorer state for prefix tokens
+            x (paddle.Tensor): The encoder feature that generates ys.
+
+        Returns:
+            tuple[paddle.Tensor, Any]: Tuple of
+                scores for next token that has a shape of `(n_vocab)`
+                and next state for ys
+
+        """
+        raise NotImplementedError
+
+    def final_score(self, state: Any) -> float:
+        """Score eos (optional).
+
+        Args:
+            state: Scorer state for prefix tokens
+
+        Returns:
+            float: final score
+
+        """
+        return 0.0
+
+
+class BatchScorerInterface(ScorerInterface):
+    """Batch scorer interface."""
+
+    def batch_init_state(self, x: paddle.Tensor) -> Any:
+        """Get an initial state for decoding (optional).
+
+        Args:
+            x (paddle.Tensor): The encoded feature tensor
+
+        Returns: initial state
+
+        """
+        return self.init_state(x)
+
+    def batch_score(
+        self, ys: paddle.Tensor, states: List[Any], xs: paddle.Tensor
+    ) -> Tuple[paddle.Tensor, List[Any]]:
+        """Score new token batch (required).
+
+        Args:
+            ys (paddle.Tensor): paddle.int64 prefix tokens (n_batch, ylen).
+            states (List[Any]): Scorer states for prefix tokens.
+            xs (paddle.Tensor):
+                The encoder feature that generates ys (n_batch, xlen, n_feat).
+
+        Returns:
+            tuple[paddle.Tensor, List[Any]]: Tuple of
+                batchfied scores for next token with shape of `(n_batch, n_vocab)`
+                and next state list for ys.
+
+        """
+        warnings.warn(
+            "{} batch score is implemented through for loop not parallelized".format(
+                self.__class__.__name__
+            )
+        )
+        scores = list()
+        outstates = list()
+        for i, (y, state, x) in enumerate(zip(ys, states, xs)):
+            score, outstate = self.score(y, state, x)
+            outstates.append(outstate)
+            scores.append(score)
+        scores = paddle.cat(scores, 0).view(ys.shape[0], -1)
+        return scores, outstates
+
+
+class PartialScorerInterface(ScorerInterface):
+    """Partial scorer interface for beam search.
+
+    The partial scorer performs scoring when non-partial scorer finished scoring,
+    and receives pre-pruned next tokens to score because it is too heavy to score
+    all the tokens.
+
+    Examples:
+         * Prefix search for connectionist-temporal-classification models
+             * :class:`espnet.nets.scorers.ctc.CTCPrefixScorer`
+
+    """
+
+    def score_partial(
+        self, y: paddle.Tensor, next_tokens: paddle.Tensor, state: Any, x: paddle.Tensor
+    ) -> Tuple[paddle.Tensor, Any]:
+        """Score new token (required).
+
+        Args:
+            y (paddle.Tensor): 1D prefix token
+            next_tokens (paddle.Tensor): paddle.int64 next token to score
+            state: decoder state for prefix tokens
+            x (paddle.Tensor): The encoder feature that generates ys
+
+        Returns:
+            tuple[paddle.Tensor, Any]:
+                Tuple of a score tensor for y that has a shape `(len(next_tokens),)`
+                and next state for ys
+
+        """
+        raise NotImplementedError
+
+
+class BatchPartialScorerInterface(BatchScorerInterface, PartialScorerInterface):
+    """Batch partial scorer interface for beam search."""
+
+    def batch_score_partial(
+        self,
+        ys: paddle.Tensor,
+        next_tokens: paddle.Tensor,
+        states: List[Any],
+        xs: paddle.Tensor,
+    ) -> Tuple[paddle.Tensor, Any]:
+        """Score new token (required).
+
+        Args:
+            ys (paddle.Tensor): paddle.int64 prefix tokens (n_batch, ylen).
+            next_tokens (paddle.Tensor): paddle.int64 tokens to score (n_batch, n_token).
+            states (List[Any]): Scorer states for prefix tokens.
+            xs (paddle.Tensor):
+                The encoder feature that generates ys (n_batch, xlen, n_feat).
+
+        Returns:
+            tuple[paddle.Tensor, Any]:
+                Tuple of a score tensor for ys that has a shape `(n_batch, n_vocab)`
+                and next states for ys
+        """
+        raise NotImplementedError

deepspeech/decoders/utils.py

@@ -0,0 +1,34 @@
+
+__all__ = ["end_detect"]
+
+def end_detect(ended_hyps, i, M=3, D_end=np.log(1 * np.exp(-10))):
+    """End detection.
+
+    described in Eq. (50) of S. Watanabe et al
+    "Hybrid CTC/Attention Architecture for End-to-End Speech Recognition"
+
+    :param ended_hyps: dict
+    :param i: int
+    :param M: int
+    :param D_end: float
+    :return: bool
+    """
+    if len(ended_hyps) == 0:
+        return False
+    count = 0
+    best_hyp = sorted(ended_hyps, key=lambda x: x["score"], reverse=True)[0]
+    for m in range(M):
+        # get ended_hyps with their length is i - m
+        hyp_length = i - m
+        hyps_same_length = [x for x in ended_hyps if len(x["yseq"]) == hyp_length]
+        if len(hyps_same_length) > 0:
+            best_hyp_same_length = sorted(
+                hyps_same_length, key=lambda x: x["score"], reverse=True
+            )[0]
+            if best_hyp_same_length["score"] - best_hyp["score"] < D_end:
+                count += 1
+
+    if count == M:
+        return True
+    else:
+        return False

deepspeech/modules/ctc.py

@@ -32,7 +32,7 @@ except Exception as e:
 __all__ = ['CTCDecoder']
 
 
-class CTCDecoder(nn.Layer):
+class CTCDecoderBase(nn.Layer):
     def __init__(self,
                  odim,
                  enc_n_units,
@@ -65,9 +65,6 @@ class CTCDecoder(nn.Layer):
             batch_average=batch_average,
             grad_norm_type=grad_norm_type)
 
-        # CTCDecoder LM Score handle
-        self._ext_scorer = None
-
     def forward(self, hs_pad, hlens, ys_pad, ys_lens):
         """Calculate CTC loss.
 
@@ -126,6 +123,13 @@ class CTCDecoder(nn.Layer):
         """
         return ctc_utils.forced_align(ctc_probs, y, blank_id)
 
+
+class CTCDecoder(CTCDecoderBase):
+    def __init__(self,*args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # CTCDecoder LM Score handle
+        self._ext_scorer = None
+
     def _decode_batch_greedy(self, probs_split, vocab_list):
         """Decode by best path for a batch of probs matrix input.
         :param probs_split: List of 2-D probability matrix, and each consists

setup.sh

@@ -69,8 +69,8 @@ fi
 # install decoders
 python3 -c "import pkg_resources; pkg_resources.require(\"swig_decoders==1.1\")"
 if [ $? != 0 ]; then
-    cd deepspeech/decoders/swig > /dev/null
+    cd deepspeech/decoders/ctcdecoder/swig > /dev/null
-    sh setup.sh
+    bash setup.sh
     cd - > /dev/null
 fi
 python3 -c "import pkg_resources; pkg_resources.require(\"swig_decoders==1.1\")"