wow_ai_mms

A simple Python package to easily use Meta's Massively Multilingual Speech (MMS) project.

The current MMS code is using subprocess to call another Python script, which is not very convenient to use, and might lead to several issues. This package is created to address those problems and to wrap up the project in an API to easily integrate it with other projects.

• Installation
• Quickstart
  • ASR
  • ASR with Alignment
  • TTS
  • LID
• API reference
• License
• Disclaimer & Credits

Installation

1. You will need ffmpeg for audio processing
2. Install wow-ai-mms from Pypi

pip install wow-ai-mms

or from source

pip install git+https://github.com/wow-ai-ml/wow-ai-mms

3. If you want to use the Alignment model:

• you will need perl to use uroman. Check the perl website for installation instructions on different platforms.
• You will need a nightly version of torchaudio:

pip install -U --pre torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118

• You might need sox as well.

4. Fairseq has not included the MMS project yet in the released PYPI version, so until the next release, you will need to install fairseq from source:

pip uninstall fairseq && pip install git+https://github.com/facebookresearch/fairseq

Quickstart

⚠️ There is an issue with fairseq when running the code in interactive environments like Jupyter notebooks.
Please use normal Python files or use the colab notebook provided above.

ASR

You will need first to download the model weights, you can find and download all the supported models from here.

from wow_ai_mms.models.asr import ASRModel

asr = ASRModel(model='/path/to/mms/model')
files = ['path/to/media_file_1', 'path/to/media_file_2']
transcriptions = asr.transcribe(files, lang='eng', align=False)
for i, transcription in enumerate(transcriptions):
    print(f">>> file {files[i]}")
    print(transcription)

ASR with Alignment

from wow_ai_mms.models.asr import ASRModel

asr = ASRModel(model='/path/to/mms/model')
files = ['path/to/media_file_1', 'path/to/media_file_2']
transcriptions = asr.transcribe(files, lang='eng', align=True)
for i, transcription in enumerate(transcriptions):
    print(f">>> file {files[i]}")
    for segment in transcription:
        print(f"{segment['start_time']} -> {segment['end_time']}: {segment['text']}")
    print("----")

Alignment model only

from wow_ai_mms.models.alignment import AlignmentModel
    
align_model = AlignmentModel()
transcriptions = align_model.align('path/to/wav_file.wav', 
                                   transcript=["segment 1", "segment 2"],
                                   lang='eng')
for transcription in transcriptions:
    for segment in transcription:
        print(f"{segment['start_time']} -> {segment['end_time']}: {segment['text']}")

TTS

from wow_ai_mms.models.tts import TTSModel

tts = TTSModel('eng')
res = tts.synthesize("This is a simple example")
tts.save(res)

LID

Coming Soon

API reference

You can check the API reference documentation for more details.

License

Since the models are released under the CC-BY-NC 4.0 license. This project is following the same License.

Disclaimer & Credits

This project is not endorsed or certified by Meta AI and is just simplifying the use of the MMS project.
All credit goes to the authors and to Meta for open sourcing the models.
Please check their paper Scaling Speech Technology to 1000+ languages and their blog post.