audio_to_text.py

import torch
from transformers import pipeline

EOS_TOKENS = [".", "!", "?"]


def transcribe_audio(audio_fpath, max_snt_len=100):
    pipe = pipeline("automatic-speech-recognition",
                    "openai/whisper-small",
                    torch_dtype=torch.float16,
                    device="cuda:0")

    pipe.model = pipe.model.to_bettertransformer()

    outputs = pipe(audio_fpath,
                   chunk_length_s=30,
                   batch_size=8,
                   return_timestamps=True)

    sentences = []
    snt_start = None
    snt = ""
    for segment in outputs["chunks"]:
        snt += f'{segment["text"]} '
        start_time, end_time = segment["timestamp"]
        if end_time is None:
            end_time = start_time + 0.1
        if not snt_start:
            snt_start = start_time
        if (
            segment["text"].strip().split()[-1][-1] in EOS_TOKENS
            or len(snt) > max_snt_len
        ):
            sentences.append(
                {"text": snt.strip(), "start": snt_start, "end": end_time}
            )
            snt_start = None
            snt = ""

    if len(snt) > 0:
        sentences.append(
            {"text": snt.strip(), "start": snt_start, "end": end_time}
        )
        snt_start = None
        snt = ""

    timestamped_text = ""
    for sentence in sentences:
        timestamped_text += (
            f'{sentence["start"]} {sentence["end"]} {sentence["text"]}\n'
        )
    return timestamped_text