Speechcatcher

This is a Python utility to interface Speechcatcher EspNet2 models. You can transcribe media files and use the utility for live transcription. All models are trained end-to-end with punctuation - the ASR model is able to output full text directly, without the need for punctuation reconstruction. Speechcatcher runs fast on CPUs and does not need a GPU to transcribe your audio.

The current focus is on German ASR. But more models will follow - stay tuned!

News

15.12.2023. New in version 0.3.1: Support for timestamps on the token level. Speechcatcher is now using espnet_streaming_decoder instead of espnet, to make dependencies leaner and to enable token timestamps with streaming models. Speechcatcher does not require a full Espnet installation anymore. It also uses a forked version of espnet_model_zoo, so that model downloads are only checked online if a local cache copy isn't available.

Installation instructions:

Install portaudio, on Mac:

brew install portaudio ffmpeg

on Linux:

sudo apt-get install portaudio19-dev python3.10-dev ffmpeg

For a system-wide and global installation, simply do:

pip3 install git+https://github.com/speechcatcher-asr/speechcatcher

Virtual environment

If you prefer an installation in a virtual environment, create one first:

virtualenv -p python3.11 speechcatcher_env

Note, if you get "-bash: virtualenv: command not found", install virtualenv through pip:

#sudo pip3 install virtualenv

Activate it:

source speechcatcher_env/bin/activate

Then install speechcatcher:

pip3 install git+https://github.com/speechcatcher-asr/speechcatcher

Run speechcatcher from the command line

After you have succesfully installed speechcatcher, you can decode any media file with:

speechcatcher media_file.mp4

or to transcribe data live from your microphone:

speechcatcher -l

All required model files are downloaded automatically and placed into a ".cache" directory.

Use speechcatcher in your Python code

To use speechcatcher in your Python script, you need to import the speechcatcher package and use the recognize function in a 'main' guarded block. Here is a complete example, that reads a 16kHz mono wav and outputs the recognized text:

from speechcatcher import speechcatcher
import numpy as np
from scipy.io import wavfile

# you need to run speechcatcher in a '__main__' guarded block:
if __name__ == '__main__':
    short_tag = 'de_streaming_transformer_xl'
    speech2text = speechcatcher.load_model(speechcatcher.tags[short_tag])

    wav_file = 'input.wav'
    rate, audio_data = wavfile.read(wav_file)
    speech = audio_data.astype(np.int16)

    print(f"Sample Rate: {rate} Hz")
    print(f"Audio Shape: {audio_data.shape}")

    # speech is a numpy array of dtype='np.int16' (16bit audio with 16kHz sampling rate)
    complete_text, paragraphs = speechcatcher.recognize(speech2text, speech, rate, quiet=True, progress=False)

    # complete_text is a string with the complete decoded text
    print(complete_text)
    
    # -> Faust. Eine Tragödie von Johann Wolfgang von Goethe. Zueignung. Ihr naht euch wieder, schwankende Gestalten...

    # paragraphs contains a list of paragraphs with additional information, such as start and end position,
    # token and token_timestamps (upper bound, in seconds)
    print(paragraphs)
    
    # -> [{'start': 0, 'end': 44.51, 'text': 'Faust. Eine Tragödie von Johann Wolfgang von Goethe. Zueignung. Ihr naht euch wieder, schwankende Gestalten...', 'tokens': ['▁F', 'aus', 't', '.', '▁Ein', 'e', '▁Tra', 'g', 'ö', 'di', 'e', '▁von', '▁Jo', 'ha', 'n', 'n', '▁Wo', 'l', 'f', 'gang', '▁von', '▁G', 'o', 'et', 'he', '.', '▁Zu', 'e', 'ig', 'n', 'ung', '.', '▁I', 'hr', '▁', 'na', 'ht', '▁euch', '▁wieder', ',', '▁sch', 'wa', 'n', 'ken', 'de', '▁Ge', 'st', 'al', 'ten', '.', ... ],
    # 'token_timestamps': [1.666, 2.333, 2.333, 3.0, 3.0, 3.0, 3.0, 3.0, 3.666, 4.333, 4.333, 4.333, 5.0, 5.0, 5.0, 5.0, 5.0, 5.666, 5.666, 5.666, 6.333, 6.333, 6.333, 7.0, 7.666, 7.666, 7.666, 7.666, 8.333, 9.666, 9.666, 9.666, 9.666, 9.666, 9.666, 10.333, 10.333, 11.0, 11.666, 11.666, 11.666, 11.666, 11.666, 12.333, 12.333, 12.333, 13.0, 13.666, 14.333, 14.333, 14.333, 14.333, 14.333, 14.333, 14.333, ... ]}, ... ]

Available models

Acoustic model	Training data (hours)	Tuda-raw test WER (without LM)	CER
de_streaming_transformer_m	13k	11.57	3.38
de_streaming_transformer_l	13k	9.65	2.76
de_streaming_transformer_xl	26k	8.5	2.44
---	---	---	---
whisper large	?	coming	soon!

Note: Tuda-de-raw results are based on raw lowercased tuda-de test utterances without the normalization step. It may not be directly comparable to regular tuda-de results.

Speechcatcher CLI parameters

Speechcatcher utility to decode speech with speechcatcher espnet models.

positional arguments:
  inputfile             Input media file

options:
  -h, --help            show this help message and exit
  -l, --live-transcription
                        Use microphone for live transcription
  -t MAX_RECORD_TIME, --max-record-time MAX_RECORD_TIME
                        Maximum record time in seconds (live transcription).
  -m MODEL, --model MODEL
                        Choose a model: de_streaming_transformer_m, de_streaming_transformer_l or de_streaming_transformer_xl
  -d DEVICE, --device DEVICE
                        Computation device. Either 'cpu' or 'cuda'. Note: Mac M1 / mps support isn't available yet.
  --lang LANGUAGE       Explicitly set language, default is empty = deduct language from model tag
  -b BEAMSIZE, --beamsize BEAMSIZE
                        Beam size for the decoder
  --quiet               No partial transcription output when transcribing a media file
  --no-progress         Show no progress bar when transcribing a media file
  --no-exception-on-overflow
                        Do not abort live recognition when encountering an input overflow.
  --save-debug-wav      Save recording to debug.wav, only applicable to live decoding
  --num-threads NUM_THREADS
                        Set number of threads used for intraop parallelism on CPU in pytorch.
  -n NUM_PROCESSES, --num-processes NUM_PROCESSES
                        Set number of processes used for processing long audio files in parallel (the input file needs to be long enough). If set to -1, use multiprocessing.cpu_count() divided by two.

Speechcatcher training

Speechcatcher models are trained by using Whisper large as a teacher model:

See speechcatcher-data for code and more info on replicating the training process.

Citation

If you use speechcatcher models in your research, for now just cite this repository:

@misc{milde2023speechcatcher,
  author = {Milde, Benjamin},
  title = {Speechcatcher},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/speechcatcher-asr/speechcatcher}},
}

the01 / speechcatcher