diff --git a/README.md b/README.md index e0769720f..468f42a61 100644 --- a/README.md +++ b/README.md @@ -9,34 +9,34 @@ English | [简体中文](README_ch.md)

-

- Quick Start - | Tutorials - | Models List - +

+ Quick Start + | Tutorials + | Models List +

- + ------------------------------------------------------------------------------------ ![License](https://img.shields.io/badge/license-Apache%202-red.svg) ![python version](https://img.shields.io/badge/python-3.7+-orange.svg) ![support os](https://img.shields.io/badge/os-linux-yellow.svg) -**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for two critical tasks in Speech - **Automatic Speech Recognition (ASR)** and **Text-To-Speech Synthesis (TTS)**, with modules involving state-of-art and influential models. +**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for two critical tasks in Speech - **Automatic Speech Recognition (ASR)** and **Text-To-Speech Synthesis (TTS)**, with modules involving state-of-art and influential models. Via the easy-to-use, efficient, flexible and scalable implementation, our vision is to empower both industrial application and academic research, including training, inference & testing module, and deployment. Besides, this toolkit also features at: - **Fast and Light-weight**: we provide a high-speed and ultra-lightweight model that is convenient for industrial deployment. -- **Rule-based Chinese frontend**: our frontend contains Text Normalization (TN) and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context. -- **Varieties of Functions that Vitalize Research**: +- **Rule-based Chinese frontend**: our frontend contains Text Normalization (TN) and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context. +- **Varieties of Functions that Vitalize Research**: - *Integration of mainstream models and datasets*: the toolkit implements modules that participate in the whole pipeline of both ASR and TTS, and uses datasets like LibriSpeech, LJSpeech, AIShell, etc. See also [model lists](#models-list) for more details. - *Support of ASR streaming and non-streaming data*: This toolkit contains non-streaming/streaming models like [DeepSpeech2](http://proceedings.mlr.press/v48/amodei16.pdf), [Transformer](https://arxiv.org/abs/1706.03762), [Conformer](https://arxiv.org/abs/2005.08100) and [U2](https://arxiv.org/pdf/2012.05481.pdf). - -Let's install PaddleSpeech with only a few lines of code! + +Let's install PaddleSpeech with only a few lines of code! >Note: The official name is still deepspeech. 2021/10/26 @@ -44,7 +44,7 @@ Let's install PaddleSpeech with only a few lines of code! # 1. Install essential libraries and paddlepaddle first. # install prerequisites sudo apt-get install -y sox pkg-config libflac-dev libogg-dev libvorbis-dev libboost-dev swig python3-dev libsndfile1 -# `pip install paddlepaddle-gpu` instead if you are using GPU. +# `pip install paddlepaddle-gpu` instead if you are using GPU. pip install paddlepaddle # 2.Then install PaddleSpeech. @@ -109,7 +109,7 @@ If you want to try more functions like training and tuning, please see [ASR gett PaddleSpeech ASR supports a lot of mainstream models, which are summarized as follow. For more information, please refer to [ASR Models](./docs/source/asr/released_model.md). @@ -125,7 +125,7 @@ The current hyperlinks redirect to [Previous Parakeet](https://github.com/Paddle - + @@ -199,7 +199,7 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model - @@ -292,11 +292,11 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
Acoustic Model Aishell2 Conv + 5 LSTM layers with only forward direction 2 Conv + 5 LSTM layers with only forward direction Ds2 Online Aishell Model
Text Frontend   + chinese-fronted
-## Tutorials +## Tutorials Normally, [Speech SoTA](https://paperswithcode.com/area/speech) gives you an overview of the hot academic topics in speech. If you want to focus on the two tasks in PaddleSpeech, you will find the following guidelines are helpful to grasp the core ideas. -The original ASR module is based on [Baidu's DeepSpeech](https://arxiv.org/abs/1412.5567) which is an independent product named [DeepSpeech](https://deepspeech.readthedocs.io). However, the toolkit aligns almost all the SoTA modules in the pipeline. Specifically, these modules are +The original ASR module is based on [Baidu's DeepSpeech](https://arxiv.org/abs/1412.5567) which is an independent product named [DeepSpeech](https://deepspeech.readthedocs.io). However, the toolkit aligns almost all the SoTA modules in the pipeline. Specifically, these modules are * [Data Prepration](docs/source/asr/data_preparation.md) * [Data Augmentation](docs/source/asr/augmentation.md) @@ -318,4 +318,3 @@ PaddleSpeech is provided under the [Apache-2.0 License](./LICENSE). ## Acknowledgement PaddleSpeech depends on a lot of open source repos. See [references](docs/source/asr/reference.md) for more information. - diff --git a/docs/source/asr/models_introduction.md b/docs/source/asr/models_introduction.md index ab2b8bac9..c99093bd6 100644 --- a/docs/source/asr/models_introduction.md +++ b/docs/source/asr/models_introduction.md @@ -13,7 +13,7 @@ In addition, the training process and the testing process are also introduced. The arcitecture of the model is shown in Fig.1.

- +
Fig.1 The Arcitecture of deepspeech2 online model

@@ -160,7 +160,7 @@ The deepspeech2 offline model is similarity to the deepspeech2 online model. The The arcitecture of the model is shown in Fig.2.

- +
Fig.2 The Arcitecture of deepspeech2 offline model

diff --git a/docs/source/asr/quick_start.md b/docs/source/asr/quick_start.md index ecce07434..da1620e90 100644 --- a/docs/source/asr/quick_start.md +++ b/docs/source/asr/quick_start.md @@ -54,7 +54,7 @@ CUDA_VISIBLE_DEVICES=0 bash local/tune.sh The grid search will print the WER (word error rate) or CER (character error rate) at each point in the hyper-parameters space, and draw the error surface optionally. A proper hyper-parameters range should include the global minima of the error surface for WER/CER, as illustrated in the following figure.

- +
An example error surface for tuning on the dev-clean set of LibriSpeech

diff --git a/docs/source/tts/models_introduction.md b/docs/source/tts/models_introduction.md index 87b65c514..b13297582 100644 --- a/docs/source/tts/models_introduction.md +++ b/docs/source/tts/models_introduction.md @@ -27,14 +27,14 @@ At present, there are two mainstream acoustic model structures. - Acoustic decoder (N Frames - > N Frames).
-
+
- Sequence to sequence acoustic model: - M Tokens - > N Frames.
-
+
### Tacotron2 @@ -54,7 +54,7 @@ At present, there are two mainstream acoustic model structures. - CBHG postprocess. - Vocoder: Griffin-Lim.
-
+
**Advantage of Tacotron:** @@ -89,7 +89,7 @@ At present, there are two mainstream acoustic model structures. - The alignment matrix of previous time is considered at the step `t` of decoder.
-
+
You can find PaddleSpeech TTS's tacotron2 with LJSpeech dataset example at [examples/ljspeech/tts0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0). @@ -118,7 +118,7 @@ Transformer TTS is a combination of Tacotron2 and Transformer. - Positional Encoding.
-
+
#### Transformer TTS @@ -138,7 +138,7 @@ Transformer TTS is a seq2seq acoustic model based on Transformer and Tacotron2. - Uniform scale position encoding may have a negative impact on input or output sequences.
-
+
**Disadvantages of Transformer TTS:** @@ -184,14 +184,14 @@ Instead of using the encoder-attention-decoder based architecture as adopted by • Can be generated in parallel (decoding time is less affected by sequence length)
-
+
#### FastPitch [FastPitch](https://arxiv.org/abs/2006.06873) follows FastSpeech. A single pitch value is predicted for every temporal location, which improves the overall quality of synthesized speech.
-
+
#### FastSpeech2 @@ -209,7 +209,7 @@ Instead of using the encoder-attention-decoder based architecture as adopted by FastSpeech2 is similar to FastPitch but introduces more variation information of speech.
-
+
You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example at [examples/csmsc/tts3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3), We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2. @@ -223,7 +223,7 @@ You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example - Describe a simple data augmentation technique that can be used early in the training to make the teacher network robust to sequential error propagation.
-
+
You can find PaddleSpeech TTS's SpeedySpeech with CSMSC dataset example at [examples/csmsc/tts2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2). @@ -289,7 +289,7 @@ You can find PaddleSpeech TTS's WaveFlow with LJSpeech dataset example at [examp - Multi-resolution STFT loss.
-
+
You can find PaddleSpeech TTS's Parallel WaveGAN with CSMSC example at [examples/csmsc/voc1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1). diff --git a/examples/other/text_frontend/README.md b/examples/other/text_frontend/README.md index 0ec4ec70f..0bf6e72dc 100644 --- a/examples/other/text_frontend/README.md +++ b/examples/other/text_frontend/README.md @@ -21,72 +21,18 @@ Run the command below to get the results of test. ``` The `avg WER` of g2p is: 0.027495061517943988 ```text - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - - ,------------------------------------------------------------------------. - | ./exp/g2p/text.g2p | - |------------------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | - |------+-----------------+-----------------------------------------------| - | bak | 9996 299181 | 290969 8198 14 14 8226 5249 | - |========================================================================| - | Sum | 9996 299181 | 290969 8198 14 14 8226 5249 | - |========================================================================| - | Mean |9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - |Median|9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | - `------------------------------------------------------------------------' - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - ,--------------------------------------------------------------------. - | ./exp/g2p/text.g2p | - |--------------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + | | # Snt # Wrd | Corr Sub Del Ins Err S.Err | |--------+-----------------+-----------------------------------------| - | bak | 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | - |====================================================================| | Sum/Avg| 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | - |====================================================================| - | Mean |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - | Median |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | `--------------------------------------------------------------------' ``` The `avg CER` of text normalization is: 0.006388318503308237 ```text - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - - ,----------------------------------------------------------------. - | ./exp/textnorm/text.tn | - |----------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | - |------+--------------+------------------------------------------| - | utt | 125 2254 | 2241 2 11 2 15 4 | - |================================================================| - | Sum | 125 2254 | 2241 2 11 2 15 4 | - |================================================================| - | Mean |125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - |Median|125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | - `----------------------------------------------------------------' - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - ,-----------------------------------------------------------------. - | ./exp/textnorm/text.tn | - |-----------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + | | # Snt # Wrd | Corr Sub Del Ins Err S.Err | |--------+--------------+-----------------------------------------| - | utt | 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | - |=================================================================| | Sum/Avg| 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | - |=================================================================| - | Mean |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - | Median |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | `-----------------------------------------------------------------' ``` diff --git a/examples/tiny/s0/README.md b/examples/tiny/s0/README.md index 7dc16dc37..11118dc45 100644 --- a/examples/tiny/s0/README.md +++ b/examples/tiny/s0/README.md @@ -37,4 +37,3 @@ ```bash bash local/export.sh ckpt_path saved_jit_model_path ``` - diff --git a/parakeet/data/batch.py b/parakeet/data/batch.py index 5e7ac3996..515074d14 100644 --- a/parakeet/data/batch.py +++ b/parakeet/data/batch.py @@ -53,8 +53,8 @@ def batch_text_id(minibatch, pad_id=0, dtype=np.int64): peek_example = minibatch[0] assert len(peek_example.shape) == 1, "text example is an 1D tensor" - lengths = [example.shape[0] for example in - minibatch] # assume (channel, n_samples) or (n_samples, ) + lengths = [example.shape[0] for example in minibatch + ] # assume (channel, n_samples) or (n_samples, ) max_len = np.max(lengths) batch = [] diff --git a/parakeet/exps/tacotron2/ljspeech.py b/parakeet/exps/tacotron2/ljspeech.py index 59c855eb6..20dc29d37 100644 --- a/parakeet/exps/tacotron2/ljspeech.py +++ b/parakeet/exps/tacotron2/ljspeech.py @@ -67,16 +67,19 @@ class LJSpeechCollector(object): # Sort by text_len in descending order texts = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(texts, text_lens), key=lambda x: x[1], reverse=True) ] mels = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(mels, text_lens), key=lambda x: x[1], reverse=True) ] mel_lens = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(mel_lens, text_lens), key=lambda x: x[1], reverse=True) ]