app.py

import os

import spaces
import torch
from transformers import pipeline, WhisperTokenizer
import torchaudio
import gradio as gr
# Please note that the below import will override whisper LANGUAGES to add bambara
# this is not the best way to do it but at least it works. for more info check the bambara_utils code
from bambara_utils import BambaraWhisperTokenizer

# Determine the appropriate device (GPU or CPU)
device = "cuda" if torch.cuda.is_available() else "cpu"

# Define the model checkpoint and language
# model_checkpoint = "oza75/whisper-bambara-asr-002" # first model
# revision = "831cd15ed74a554caac9f304cf50dc773841ba1b"
# model_checkpoint = "oza75/whisper-bambara-asr-005"
# revision = "6a92cd0f19985d12739c2f6864607627115e015d" # first good checkpoint for bambara

#revision = "fb69a5750182933868397543366dbb63747cf40c" # this only translate in english
#revision = "129f9e68ead6cc854e7754b737b93aa78e0e61e1" # support transcription and translation
#revision = "cb8e351b35d6dc524066679d9646f4a947300b27"
#revision = "5f143f6070b64412a44fea08e912e1b7312e9ae9" # this checkpoint support both task without overfitting

#model_checkpoint = "oza75/whisper-bambara-asr-006"
#revision = "96535debb4ce0b7af7c9c186d09d088825f63840"
#revision = "4549778c08f29ed2e033cc9a497a187488b6bf56"

# model_checkpoint = "oza75/bm-whisper-02"
# revision = "06e81aa0214f6d07d3d787b367e3e8357b171549"

# language = "bambara"
# language = "icelandic" # we use icelandic as the model was trained to replace the icelandic with bambara.

#model_checkpoint = "oza75/bm-whisper-from-swa-02"
revision = None
#language = "swahili"

#model_checkpoint = "oza75/bm-whisper-large-turbo-v4"
#model_checkpoint = "oza75/bm-whisper-large-v3-base"
# model_checkpoint = "oza75/bm-whisper-large-v3-sft-3"
model_checkpoint = "djelia/bm-whisper-large-v2-lora-merged"
# language = "sundanese"

# Load the custom tokenizer designed for Bambara and the ASR model
#tokenizer = BambaraWhisperTokenizer.from_pretrained(model_checkpoint, language=language, device=device)
tokenizer = WhisperTokenizer.from_pretrained(model_checkpoint, device=device)
pipe = pipeline("automatic-speech-recognition", model=model_checkpoint, tokenizer=tokenizer, device=device, revision=revision)

LANGUAGES = {"bambara": "french", "french": "french", "english": "english"}

def resample_audio(audio_path, target_sample_rate=16000):
    """
    Converts the audio file to the target sampling rate (16000 Hz).
    
    Args:
        audio_path (str): Path to the audio file.
        target_sample_rate (int): The desired sample rate.

    Returns:
        A tensor containing the resampled audio data and the target sample rate.
    """
    waveform, original_sample_rate = torchaudio.load(audio_path)
    
    if original_sample_rate != target_sample_rate:
        resampler = torchaudio.transforms.Resample(orig_freq=original_sample_rate, new_freq=target_sample_rate)
        waveform = resampler(waveform)
    
    return waveform, target_sample_rate

@spaces.GPU()
def transcribe(audio, task_type, language):
    """
    Transcribes the provided audio file into text using the configured ASR pipeline.

    Args:
        audio: The path to the audio file to transcribe.

    Returns:
        A string representing the transcribed text.
    """
    # Convert the audio to 16000 Hz
    waveform, sample_rate = resample_audio(audio)
    
    language = LANGUAGES[language]
    
    # Use the pipeline to perform transcription
    sample = {"array": waveform.squeeze().numpy(), "sampling_rate": sample_rate}
    text = pipe(sample, generate_kwargs={
        "task": task_type,
        "num_beams": 1, 
        "early_stopping": True, 
        "language": language
    })["text"]
    
    return text

def get_wav_files(directory):
    """
    Returns a list of absolute paths to all .wav files in the specified directory.

    Args:
        directory (str): The directory to search for .wav files.

    Returns:
        list: A list of absolute paths to the .wav files.
    """
    # List all files in the directory
    files = os.listdir(directory)
    # Filter for .wav files and create absolute paths
    wav_files = [os.path.abspath(os.path.join(directory, file)) for file in files if file.endswith('.wav')]
    wav_files = [[f, "transcribe", "bambara"] for f in wav_files]

    return wav_files

def main():
    # Get a list of all .wav files in the examples directory
    example_files = get_wav_files("./examples")

    # Setup Gradio interface
    iface = gr.Interface(
        fn=transcribe,
        inputs=[
            gr.Audio(type="filepath", value=example_files[0][0]),
            gr.Radio(choices=["transcribe", "translate"], label="Task Type", value="transcribe"),
            gr.Dropdown(choices=LANGUAGES.keys(), label="Language", value="bambara"),
            
        ],
        outputs="text",
        title="Bambara Automatic Speech Recognition",
        description="Realtime demo for Bambara speech recognition based on a fine-tuning of the Whisper model.",
        examples=example_files,
        cache_examples="lazy",
    )

    # Launch the interface
    iface.launch(share=False)


if __name__ == "__main__":
    main()