app.py

import gradio as gr
from transformers import pipeline, Wav2Vec2ProcessorWithLM
import os

def transcribe(audio, model_id, model_revison):
    if model_revison != "main":
        # load processor
        p = Wav2Vec2ProcessorWithLM.from_pretrained(model_id, revision=model_revison)
        # load eval pipeline
        asr = pipeline("automatic-speech-recognition", model=model_id, tokenizer=p.tokenizer, feature_extractor=p.feature_extractor, decoder=p.decoder, token=os.getenv('HF_TOKEN'))
    else:
        # load eval pipeline
        asr = pipeline("automatic-speech-recognition", model=model_id, token=os.getenv('HF_TOKEN'))

    text = asr(audio)["text"]
    return text

asr_app = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources=["upload", "microphone"], type="filepath"),
        gr.Dropdown(
            [
                "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0",
                "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0",
                "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0"
            ]
        ),
        gr.Radio(["main","lm"])
    ],
    examples=[
        ["./examples/CV/hausa/common_voice_ha_32885169.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0", "lm"],
        ["./examples/CV/hausa/common_voice_ha_29417456.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0", "lm"],
        ["./examples/CV/hausa/common_voice_ha_28554827.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0", "main"],
        ["./examples/CV/igbo/common_voice_ig_31594237.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0", "lm"],
        ["./examples/CV/igbo/common_voice_ig_30710992.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0", "lm"],
        ["./examples/CV/igbo/common_voice_ig_30692048.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0", "main"],
        ["./examples/CV/yoruba/common_voice_yo_36914062.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0", "lm"],
        ["./examples/CV/yoruba/common_voice_yo_36841367.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0", "lm"],
        ["./examples/CV/yoruba/common_voice_yo_36526475.wav", "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0", "main"]
    ],
    outputs="text",
    title="NaijaVoices ASR",
    description="Realtime demo for Hausa, Igbo and Yoruba speech recognition using a fine-tuned Wav2Vec2-XLS-R 1B model.",
)

asr_app.launch()