app.py

import gradio as gr
import soundfile as sf
import numpy as np
import torch, torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

MODEL_IS="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"
MODEL_FO="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h"

torch.random.manual_seed(0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model_is = Wav2Vec2ForCTC.from_pretrained(MODEL_IS).to(device)
processor_is = Wav2Vec2Processor.from_pretrained(MODEL_IS) 
model_fo = Wav2Vec2ForCTC.from_pretrained(MODEL_FO).to(device)
processor_fo = Wav2Vec2Processor.from_pretrained(MODEL_FO) 

def readwav(a_f):
    wav, sr = sf.read(a_f, dtype=np.float32)
    if len(wav.shape) == 2:
        wav = wav.mean(1)
    if sr != 16000:
        wlen = int(wav.shape[0] / sr * 16000)
        wav = signal.resample(wav, wlen)
    return wav

def recis(audio_file):
    wav = readwav(audio_file)
    with torch.inference_mode():
        input_values = processor_is(wav,sampling_rate=16000).input_values[0]
        input_values = torch.tensor(input_values, device=device).unsqueeze(0)
        logits = model_is(input_values).logits
        pred_ids = torch.argmax(logits, dim=-1)
        xcp = processor_is.batch_decode(pred_ids)
        return xcp

def recfo(audio_file):
    wav = readwav(audio_file)
    with torch.inference_mode():
        input_values = processor_fo(wav,sampling_rate=16000).input_values[0]
        input_values = torch.tensor(input_values, device=device).unsqueeze(0)
        logits = model_fo(input_values).logits
        pred_ids = torch.argmax(logits, dim=-1)
        xcp = processor_fo.batch_decode(pred_ids)
        return xcp

bl = gr.Blocks()
with bl:

    gr.Markdown(
        """
    # W2V2 speech recognition
    Upload a file for recognition with 
    https://huggingface.co/carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h 
    or https://huggingface.co/carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h  

    - For some reason, the huggingface 'Hosted inference API' on the model page does not work, but this demo does.  
    - There is no language model (yet), so it can generate non-words.
    """
    )

    with gr.Tabs():
        with gr.TabItem("Icelandic"):
            with gr.Row():
                audio_file = gr.Audio(type="filepath")
                text_output = gr.Textbox()
            text_button = gr.Button("Recognise")
            text_button.click(recis, inputs=audio_file, outputs=text_output)
        with gr.TabItem("Faroese"):
            with gr.Row():
                audio_file = gr.Audio(type="filepath")
                text_output = gr.Textbox()
            text_button = gr.Button("Recognise")
            text_button.click(recfo, inputs=audio_file, outputs=text_output)

bl.launch()