Data contains transcripts file and audio files. The format of transcription
file is csv in the form of <transcript>,<set_id>,<speaker_id>,<file path> for
each transcript. If speaker ID is not available, any word(e.g None) can be
used to normalize the features by complete dataset.
To change the vocabulary, you need to edit the python code in vocab.py
Install tensorflow, python_speech_features and openfst(optional) first before
running, and generate FST if set in command line options.
usage: deepsphinx-train [-h] [--nfilt NFILT] [--keep-prob KEEP_PROB]
[--max-output-len MAX_OUTPUT_LEN]
[--rnn-size RNN_SIZE] [--num-layers NUM_LAYERS]
[--num-decoding-layers NUM_DECODING_LAYERS]
[--batch-size BATCH_SIZE]
[--learning-rate LEARNING_RATE]
[--num-epochs NUM_EPOCHS] [--beam-width BEAM_WIDTH]
[--learning-rate-decay LEARNING_RATE_DECAY]
[--min-learning-rate MIN_LEARNING_RATE]
[--display-step DISPLAY_STEP] [--data-dir DATA_DIR]
[--job-dir JOB_DIR]
[--checkpoint-path CHECKPOINT_PATH]
[--best-n-inference BEST_N_INFERENCE]
[--eval-only EVAL_ONLY] [--fst-path FST_PATH]
[--cutoff-range CUTOFF_RANGE]
[--use-inference-lm [USE_INFERENCE_LM]]
[--nouse-inference-lm] [--use-train-lm [USE_TRAIN_LM]]
[--nouse-train-lm]
[--use-conv-feat-att [USE_CONV_FEAT_ATT]]
[--nouse-conv-feat-att] [--trans-file TRANS_FILE]
optional arguments:
-h, --help show this help message and exit
--nfilt NFILT Number of filters in mel specta
--keep-prob KEEP_PROB
Keep probability for dropout
--max-output-len MAX_OUTPUT_LEN
Maximum output length for Beam Search
--rnn-size RNN_SIZE Size of LSTM
--num-layers NUM_LAYERS
Number of LSTM Layers for encoding layer
--num-decoding-layers NUM_DECODING_LAYERS
Number of LSTM Layers for decoding layer
--batch-size BATCH_SIZE
Batch size of data
--learning-rate LEARNING_RATE
Learning rate
--num-epochs NUM_EPOCHS
Number of epochs
--beam-width BEAM_WIDTH
Beam width. Must be lesser than vocab size
--learning-rate-decay LEARNING_RATE_DECAY
Learning rate decay
--min-learning-rate MIN_LEARNING_RATE
Minimum learning rate
--display-step DISPLAY_STEP
Check training loss after every display-step batches
--data-dir DATA_DIR Directory of data
--job-dir JOB_DIR Directory in which summary and checkpoint is stored
--checkpoint-path CHECKPOINT_PATH
Load a trained model
--best-n-inference BEST_N_INFERENCE
Take best of n for beam search
--eval-only EVAL_ONLY
Only evaluate. --checkpoint-path is required
--fst-path FST_PATH Path of language FST
--cutoff-range CUTOFF_RANGE
Attention cutoff from previous mean range
--use-inference-lm [USE_INFERENCE_LM]
Use LM during inference
--nouse-inference-lm
--use-train-lm [USE_TRAIN_LM]
Use LM during training
--nouse-train-lm
--use-conv-feat-att [USE_CONV_FEAT_ATT]
Use convolutional features for attention
--nouse-conv-feat-att
--trans-file TRANS_FILE
Path of transcription file
The model also supports using FST based LM. See file make_fst for commands
for generating Language model FST.
There is a minimal API present for prediction of sequence from audio files. As it is not fully complete, it is recommended to read it before using it. The usage is intended to be something like this:
from deepsphinx.api import Predict
ds = Predict('path/to/checkpoint').predict('path/to/audio')
There is a tool for inference on single file. We normalize the features by file, and so the accuracy might be lower as for training we normalize by speaker.
Run this command for inference:
deepsphinx-infer --checkpoint-path <ckpt> --audio-file <audio-file>
To get all the options run deepsphinx-infer. Options need to be the same as
that used for training session for the checkpoint
https://s3.amazonaws.com/deepsphinx/batch-21937.data-00000-of-00001 https://s3.amazonaws.com/deepsphinx/batch-21937.index