app.py

import gradio as gr
from transformers import Wav2Vec2Processor, AutoModelForCTC
from transformers import Wav2Vec2ProcessorWithLM
import torch
from pyctcdecode import build_ctcdecoder
import librosa
import logging
# This hides the 'Exception ignored in' messages which are usually harmless during shutdown
logging.getLogger("asyncio").setLevel(logging.CRITICAL)
# Replace with your specific model path
MODEL_ID = "eleferrand/w2v-Morisyen" 

# Initialize the ASR pipeline
if torch.cuda.is_available():
    device="cuda"
else:
    device="cpu"
    
lm_model= "Morisyen.arpa"
path_checkpoint = MODEL_ID

model = AutoModelForCTC.from_pretrained(path_checkpoint).to(device)
processor = Wav2Vec2Processor.from_pretrained(path_checkpoint)

vocab = processor.tokenizer.get_vocab()

vocab[' '] = vocab['|']
del vocab[' ']

sorted_dict = {k.lower(): v for k, v in sorted(vocab.items(), key=lambda item: item[1])}
print(sorted_dict)

decoder = build_ctcdecoder(
    list(sorted_dict.keys()),
    lm_model,
    alpha = 0.5,
    beta = 1.5
)

processor_with_lm = Wav2Vec2ProcessorWithLM(
    feature_extractor=processor.feature_extractor,
    tokenizer=processor.tokenizer,
    decoder=decoder
)

def transcribe(audio_path, request: gr.Request):
    if audio_path is None:
        return "No audio recorded", ""
    is_api = request.headers.get("sec-ch-ua") is None
    w, sr = librosa.load(audio_path, sr=16000)
    entry = {"input_values" : processor(w, sampling_rate=sr).input_values[0]}
    entry["input_length"] = len(entry["input_values"])

    input_dict = processor_with_lm(entry["input_values"], return_tensors="pt",sampling_rate=16000, padding=True)
    logits = model(input_dict.input_values.to(device)).logits
    transc = processor_with_lm.decode(logits[0].cpu().detach().numpy()).text
    if is_api==False:
        direction = ""
        if "dwat" in transc:
            direction = " ale a dwat..."
        elif "gos" in transc:
            direction = "ale a gos..."
        elif "dwa" in transc:
            direction = "ale tout dwa"
        elif "se bon" in transc:
            direction = "end"
        else:
            direction = transc
        return direction, transc
    else:
        return transc, "something"

# Define the interface

demo = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(type="filepath"),
    outputs=[gr.Text(), gr.Text()],
    api_name="transcribe" 
)

# Use this to prevent background threads from hanging on exit
demo.launch(show_error=True)