src/runners/audio_transcriber.py

import tempfile

import bentoml
import ffmpeg
import numpy as np
import torch
from faster_whisper import WhisperModel
from transformers import pipeline


class AudioTranscriber(bentoml.Runnable):
    SUPPORTED_RESOURCES = ("nvidia.com/gpu", "cpu")
    SUPPORTS_CPU_MULTI_THREADING = True

    SAMPLE_RATE = 16000

    def __init__(self):
        self.faster_model = WhisperModel("base")
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        # self.model.to(self.device)

    @bentoml.Runnable.method(batchable=False)
    def transcribe_audio_faster(self, temp_file_path):
        segments, info = self.faster_model.transcribe(temp_file_path)
        transcription = []
        segment_info = []
        for segment in segments:
            print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
            transcription.append(segment.text)
            segment_info.append((segment.start, segment.end))

        return transcription, info, segment_info

    @bentoml.Runnable.method(batchable=False)
    def transcribe_audio(self, file):
        with tempfile.NamedTemporaryFile(delete=False) as temp_file:
            temp_file.write(file.read())
            temp_file_path = temp_file.name

        # Use the temporary file path as input for ffmpeg.input()

        try:
            # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
            # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
            out, _ = (
                ffmpeg.input(temp_file_path, threads=0)
                .output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=self.SAMPLE_RATE)
                .run(cmd="ffmpeg", capture_stdout=True, capture_stderr=True)
            )
        except ffmpeg.Error as e:
            raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}")

        input_features = np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0

        print("Input_features", type(input_features))

        pipe = pipeline("automatic-speech-recognition",
                        model=self.model, tokenizer=self.tokenizer, feature_extractor=self.extractor,
                        device=self.device)

        result = self.get_long_transcription_whisper(input_features, pipe)

        return result

    @staticmethod
    def get_long_transcription_whisper(input_features, pipe, return_timestamps=True,
                                       chunk_length_s=10, stride_length_s=2):
        """Get the transcription of a long audio file using the Whisper model
        input_feature: numpy.ndarray
        """

        return pipe(input_features, return_timestamps=return_timestamps,
                    chunk_length_s=chunk_length_s, stride_length_s=stride_length_s)