inference.py

from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torchaudio
import torch


repo_id = "eddiegulay/wav2vec2-large-xlsr-mvc-swahili"

model = Wav2Vec2ForCTC.from_pretrained(repo_id)
processor = Wav2Vec2Processor.from_pretrained(repo_id)



def transcribe(audio_path):
  # Load the audio file
  audio_input, sample_rate = torchaudio.load(audio_path)
  target_sample_rate = 16000
  audio_input = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sample_rate)(audio_input)

  # Preprocess the audio data
  input_dict = processor(audio_input[0], return_tensors="pt", padding=True, sampling_rate=16000)

  # Perform inference and transcribe
  logits = model(input_dict.input_values).logits
  pred_ids = torch.argmax(logits, dim=-1)[0]
  transcription = processor.decode(pred_ids)

  return transcription


# transcribe sample audio
transcribe("download.wav")