planning: Ichigo Transcription

[dan-menlo]

Goal

• Ichigo Demo should have transcription of the audio message
• Likely driven by Eng
• Data Storage is easier (i.e. can train over it)
• Whisper Encoder is already in project (i.e. use Decoder)
• Will not affect latency as this is post-processing

[tikikun]

@nguyenhoangthuan99 you can pick this up if you like, just extract embedding from encoder and forward it to whisper transcription

[jrohsc]

Hi, how can I do the transcription on a colab notebook? It seems like whenever I give a question audio, it only generates the answer to the question.

[PodsAreAllYouNeed]

Hi, how can I do the transcription on a colab notebook? It seems like whenever I give a question audio, it only generates the answer to the question.

I've prepared a colab demo with transcription example here: https://colab.research.google.com/drive/1req3ByqKS1vVPF_iGD1sNE2DzvMo7Jd0?usp=sharing

The relevant function for transcription is this:

def audio_to_text(audio_path, target_bandwidth=1.5, device=device):
    vq_model.ensure_whisper(device)
    wav, sr = torchaudio.load(audio_path)
    if sr != 16000:
        wav = torchaudio.functional.resample(wav, sr, 16000)
    with torch.no_grad():
        codes = vq_model.encode_audio(wav.to(device))
        transcript = vq_model.decode_text(codes[0]) 
    return f'{transcript[0].text}'