app.py

import gradio as gr
from transformers import pipeline
import os

def transcribe(audio, language):
    model_map = {
        "hausa": "asr-africa/wav2vec2-xls-r-1b-naijavoices-hausa-500hr-v0",
        "igbo": "asr-africa/wav2vec2-xls-r-1b-naijavoices-igbo-500hr-v0",
        "yoruba": "asr-africa/wav2vec2-xls-r-1b-naijavoices-yoruba-500hr-v0",
        "zulu": "asr-africa/W2V2-Bert_nchlt_speech_corpus_Fleurs_ZULU_63hr_v1",
        "xhosa": "asr-africa/wav2vec2_xls_r_300m_nchlt_speech_corpus_Fleurs_XHOSA_63hr_v1",
        "afrikaans": "asr-africa/mms-1B_all_nchlt_speech_corpus_Fleurs_CV_AFRIKAANS_57hr_v1",
        "bemba": "asr-africa/whisper_BIG-C_BEMBA_189hr_v1",
        "shona": "asr-africa/W2V2_Bert_Afrivoice_FLEURS_Shona_100hr_v1",
        "luganda": "asr-africa/whisper-small-CV-Fleurs-lg-313hrs-v1",
        "swahili": "asr-africa/wav2vec2-xls-r-300m-CV_Fleurs_AMMI_ALFFA-sw-400hrs-v1",
        "lingala": "asr-africa/wav2vec2-xls-r-300m-Fleurs_AMMI_AFRIVOICE_LRSC-ln-109hrs-v2",
        "amharic": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-amh-200hrs-v1",
        "kinyarwanda": "asr-africa/facebook-mms-1b-all-common_voice_fleurs-rw-100hrs-v1",
        "oromo": "asr-africa/mms-1b-all-Sagalee-orm-85hrs-4",
        "akan": "asr-africa/wav2vec2-xls-r-akan-100-hours",
        "ewe": "asr-africa/wav2vec2-xls-r-ewe-100-hours",
        "wolof": "asr-africa/w2v2-bert-Wolof-20-hours-Google-Fleurs-ALF-dataset",
        "bambara": "asr-africa/mms-bambara-50-hours-mixed-bambara-dataset",
    }

    # load eval pipeline
    asr = pipeline("automatic-speech-recognition", model=model_map[language], device=0, 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(
            [
                "hausa",
                "igbo",
                "yoruba",
                "zulu",
                "xhosa",
                "afrikaans",
                "bemba",
                "shona",
                "luganda",
                "swahili",
                "lingala",
                "amharic",
                "kinyarwanda",
                "oromo",
                "akan",
                "ewe",
                "wolof",
                "bambara",
            ]
        ),
    ],
    outputs="text",
    title="ASR Africa",
    description="This space serves as a realtime demo for automatic speech recognition models developed by Mak-CAD under the auspicies of Gates Foundation for 18 African languages using open source data.\
    \nWe would appreciate your feedback on these models, you can share your feedback via this form https://forms.gle/RbzpwBFbC6Lcx5V78 :)"
)

asr_app.launch()