| import torch |
| import torchaudio |
| from pathlib import Path |
| import soundfile as sf |
| from typing import Any |
|
|
|
|
| from config import TARGET_SR, SUPPORTED_EXTS |
|
|
|
|
| def transcribe_file(path: str | Path, pipe: Any) -> str: |
| """ |
| Transcribe an audio file to text using a given ASR pipeline. |
| |
| Args: |
| path: Path or string pointing to an audio file. |
| asr_pipeline: A Hugging Face transformers pipeline object for |
| automatic-speech-recognition. Should accept a numpy |
| array and return a dict with key 'text'. |
| |
| Returns: |
| The transcribed text as returned by the pipeline. |
| |
| Raises: |
| ValueError: If loading or decoding the audio fails. |
| """ |
| speech = load_resample(path) |
| return pipe(speech.numpy())["text"] |
|
|
|
|
| def load_resample(path: str | Path, target_sr: int = TARGET_SR) -> torch.Tensor: |
| """ |
| Load an audio file and resample it to the target sample rate, returning |
| a mono torch.Tensor. |
| |
| Args: |
| path: Path or string pointing to an audio file. |
| target_sr: Desired sample rate (in Hz). Defaults to TARGET_SR from config. |
| |
| Returns: |
| A 1-D torch.Tensor of dtype float32 sampled at target_sr. |
| |
| Raises: |
| ValueError: If the file extension is not in SUPPORTED_EXTS. |
| ValueError: If the audio file cannot be decoded. |
| """ |
| ext = Path(path).suffix.lower() |
| if ext not in SUPPORTED_EXTS: |
| raise ValueError( |
| f"Unsupported file-type “{ext or 'unknown'}”. Please upload WAV, FLAC, MP3, OGG/Opus or M4A." |
| ) |
|
|
| try: |
| speech, sr = sf.read(str(path)) |
| except RuntimeError as exc: |
| raise ValueError( |
| "Couldn't decode the audio file - maybe it's corrupted or in an uncommon codec." |
| ) from exc |
|
|
| speech = torch.tensor(speech).float() |
| if speech.ndim == 2: |
| speech = speech.mean(dim=1) |
| if sr != target_sr: |
| speech = torchaudio.functional.resample( |
| speech, orig_freq=sr, new_freq=target_sr |
| ) |
| return speech |
|
|
