Adapt for DeepSpeech 0.6.0

[BoneGoat]

I'm trying to adapt this code for the new release of DeepSpeech. After some minor modifications in align/wavTranscriber.py, mostly CreateModel and enableDecoderWithLM, I'm running into the following error:

bin/align.sh --output-max-cer 15 --loglevel 10 --audio data/test1/audio.wav --script data/test1/transcript.txt --aligned data/test1/aligned.json --tlog data/test1/log.txt --stt-model-dir models/en/deepspeech-0.6.0-models --alphabet models/en/alphabet.txt
DEBUG:root:Start
DEBUG:root:Loading alphabet from "models/en/alphabet.txt"...
DEBUG:root:Looking for model files in "models/en/deepspeech-0.6.0-models"...
DEBUG:root:Loading acoustic model from "models/en/deepspeech-0.6.0-models/output_graph.pb", alphabet from "models/en/alphabet.txt" and language model from "data/test1/transcript.txt.lm"...
DEBUG:root:Transcribing VAD segments...
VAD splitting: 3464it [00:01, 2274.26it/s]
Transcribing:   0%|                                                                                                                                                                                                  | 0/3464 [00:00<?, ?it/s]DEBUG:root:Process 28330: Loaded models
TensorFlow: v1.14.0-21-ge77504ac6b
DeepSpeech: v0.6.0-0-g6d43e21
Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage.
2019-12-19 12:58:48.519864: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Error: Trie file version mismatch (4 instead of expected 5). Update your trie file.
libc++abi.dylib: terminating with uncaught exception of type int

pip list
Package      Version
------------ -------
deepspeech   0.6.0
numpy        1.17.4
pip          19.0.3
pydub        0.23.1
setuptools   40.8.0
six          1.13.0
sox          1.3.7
textdistance 4.1.5
tqdm         4.40.2
webrtcvad    2.0.10

I have replaced generate_trie with the 0.6.0 version from native client. I can get a bit further by not generating specific LMs but then I run into this problem:

bin/align.sh --output-max-cer 15 --loglevel 10 --audio data/test1/audio.wav --script data/test1/transcript.txt --aligned data/test1/aligned.json --tlog data/test1/log.txt --stt-model-dir models/en/deepspeech-0.6.0-models --alphabet models/en/alphabet.txt --stt-no-own-lm
DEBUG:root:Start
DEBUG:root:Loading alphabet from "models/en/alphabet.txt"...
DEBUG:root:Looking for model files in "models/en/deepspeech-0.6.0-models"...
DEBUG:root:Loading acoustic model from "models/en/deepspeech-0.6.0-models/output_graph.pb", alphabet from "models/en/alphabet.txt", trie from "models/en/deepspeech-0.6.0-models/trie" and language model from "models/en/deepspeech-0.6.0-models/lm.binary"...
DEBUG:root:Transcribing VAD segments...
VAD splitting: 3464it [00:01, 2264.24it/s]
Transcribing:   0%|                                                                                                                                                                                                  | 0/3464 [00:00<?, ?it/s]DEBUG:root:Process 31650: Loaded models
TensorFlow: v1.14.0-21-ge77504ac6b
DeepSpeech: v0.6.0-0-g6d43e21
Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage.
2019-12-19 14:19:34.495576: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
DEBUG:root:Process 31650: Transcribing...
DEBUG:root:Process 31650: Transcribing...
multiprocessing.pool.RemoteTraceback:
"""
Traceback (most recent call last):
  File "/usr/local/Cellar/python/3.7.4_1/Frameworks/Python.framework/Versions/3.7/lib/python3.7/multiprocessing/pool.py", line 121, in worker
    result = (True, func(*args, **kwds))
  File "/Users/tobias/dev/git/DSAlign/align/align.py", line 74, in stt
    transcript = wavTranscriber.stt(model, audio, sample_rate)
  File "/Users/tobias/dev/git/DSAlign/align/wavTranscriber.py", line 40, in stt
    output = ds.stt(audio, fs)
  File "/Users/tobias/dev/git/DSAlign/venv/lib/python3.7/site-packages/deepspeech/__init__.py", line 93, in stt
    return deepspeech.impl.SpeechToText(self._impl, *args, **kwargs)
TypeError: SpeechToText() takes at most 2 arguments (3 given)
"""

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/Users/tobias/dev/git/DSAlign/align/align.py", line 672, in <module>
DEBUG:root:Process 31650: Transcribing...
    main()
  File "/Users/tobias/dev/git/DSAlign/align/align.py", line 631, in main
    for time_start, time_end, segment_transcript in transcripts:
  File "/Users/tobias/dev/git/DSAlign/venv/lib/python3.7/site-packages/tqdm/std.py", line 1102, in __iter__
    for obj in iterable:
  File "/usr/local/Cellar/python/3.7.4_1/Frameworks/Python.framework/Versions/3.7/lib/python3.7/multiprocessing/pool.py", line 748, in next
    raise value
DEBUG:root:Process 31650: Transcribing...
TypeError: SpeechToText() takes at most 2 arguments (3 given)
DEBUG:root:Process 31650: Transcribing...
Transcribing:   0%|

I'm sure these are just minor problems and adapting DSAlign to DeepSpeech 0.6.0 won't be that difficult.

[BoneGoat]

Fixed the "SpeechToText() takes at most 2 arguments (3 given)" problem by removing sample rate from the ds.tts() call. Sample rate is now inferred from the model.

[tilmankamp]

Thanks for your help! Could you add/replace deepspeech==0.6.0 in the requirements.txt file and put this fixes up as a PR?

[BoneGoat]

Got passed the "Trie file version mismatch (4 instead of expected 5). Update your trie file." error by using v0.6.0 of the native client from GitHub extracted into dependencies/deepspeech. taskcluster.py is supposed to fix this but taskcluster.net seems deprecated. taskcluster.py needs refactoring to pull files from GitHub.

[BoneGoat]

Thanks for merging and for this wonderful lib.

[tilmankamp]

https://github.com/mozilla/DSAlign/commit/6dcfd4dd4a8146d23c0c784a960c1557b6e803bb should have fixed the DeepSpeech dependency in case of individual language models.