fix libri ds2 scripts; add ngram and spm doc

examples/librispeech/s0/conf/deepspeech2.yaml

@@ -10,9 +10,9 @@ data:
   min_input_len: 0.0
   max_input_len: 27.0 # second
   min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+  max_output_len: .inf
+  min_output_input_ratio: 0.00
+  max_output_input_ratio: .inf
   specgram_type: linear
   target_sample_rate: 16000
   max_freq: None
@@ -41,7 +41,7 @@ training:
   lr: 1e-3
   lr_decay: 0.83
   weight_decay: 1e-06
-  global_grad_clip: 5.0
+  global_grad_clip: 3.0
   log_interval: 100
 
 decoding:

examples/librispeech/s0/local/data.sh

@@ -61,13 +61,13 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     num_workers=$(nproc)
     python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
     --manifest_path="data/manifest.train.raw" \
-    --num_samples=-1 \
+    --num_samples=2000 \
     --specgram_type="linear" \
     --delta_delta=false \
     --sample_rate=16000 \
     --stride_ms=10.0 \
     --window_ms=20.0 \
-    --use_dB_normalization=False \
+    --use_dB_normalization=True \
     --num_workers=${num_workers} \
     --output_path="data/mean_std.json"
 

examples/ngram_lm/README.md

@@ -0,0 +1,7 @@
+# Ngram LM
+
+Train chinese chararctor ngram lm by [kenlm](https://github.com/kpu/kenlm).
+
+```
+bash run.sh
+```

examples/spm/README.md

@@ -1,4 +1,6 @@
-# SPM demo
+# [SentencePiece Model](https://github.com/google/sentencepiece)
+
+Train a `spm` model for English tokenizer.
 
 ```
 bash run.sh