allandclive

/

whisper-small-luganda

like 2

It works pretty well.

Language lg is not recognized, it works without using language option.

Script I am using:

#!/home/data1/protected/venv/bin/python

import sys
from transformers import pipeline
from pydub import AudioSegment
from pathlib import Path
import io
import torch

def transcribe_audio_in_memory(audio_path, pipe, chunk_length_ms=30_000):
    """
    Transcribe audio (any length) using Whisper Luganda model entirely in memory.
    Splits into chunks internally to avoid long-form issues.
    Returns full transcription as string.
    """
    audio = AudioSegment.from_file(audio_path)

    transcript = ""
    # Process audio in chunks
    for i in range(0, len(audio), chunk_length_ms):
        chunk = audio[i:i+chunk_length_ms]
        # Export chunk to a bytes buffer
        buf = io.BytesIO()
        chunk.export(buf, format="wav")
        buf.seek(0)
        # Transcribe directly from bytes buffer
        result = pipe(buf, generate_kwargs={"task": "transcribe"})
        transcript += result["text"] + " "

    return transcript.strip()


def save_transcription(audio_path, transcription):
    """
    Save transcription to a .txt file using the same base name as the audio file.
    """
    base_name = os.path.splitext(audio_path)[0]
    txt_file = base_name + ".txt"
    with open(txt_file, "w", encoding="utf-8") as f:
        f.write(transcription)
    print(f"Transcription saved to: {txt_file}")


if __name__ == "__main__":
    if len(sys.argv) < 2:
        print("Usage: python transcribe_multiple_in_memory.py <folder_or_audio_file1> [<audio_file2> ...]")
        sys.exit(1)

    # Initialize Whisper pipeline once
    pipe = pipeline(
        "automatic-speech-recognition",
        model="allandclive/whisper-small-luganda",
        device=0  # change to -1 for CPU
    )

    input_paths = []
    for arg in sys.argv[1:]:
        p = Path(arg)
        if p.is_file():
            input_paths.append(p)
        elif p.is_dir():
            input_paths.extend(p.glob("*.ogg"))
            input_paths.extend(p.glob("*.wav"))
            input_paths.extend(p.glob("*.mp3"))
        else:
            print(f"Skipping invalid path: {arg}")

    if not input_paths:
        print("No valid audio files found.")
        sys.exit(1)

    # Process each file
    for audio_file in input_paths:
        print(f"Processing: {audio_file}")
        transcription = transcribe_audio_in_memory(str(audio_file), pipe)
        save_transcription(str(audio_file), transcription)