app.py

import gradio as gr
import soundfile as sf
import numpy as np
import torch, torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from datasets import load_dataset, Audio
import matplotlib.pyplot as plt


MODEL_NAME="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"

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

model = Wav2Vec2ForCTC.from_pretrained(MODEL_NAME).to(device)
processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME) # do i need this? can't remember


#ds = load_dataset("language-and-voice-lab/samromur_asr",split='train',streaming=True)
#ds = load_dataset("language-and-voice-lab/samromur_asr",split='test')
#ds = ds.cast_column("audio", Audio(sampling_rate=16_000))

def show_ex(exnum):
    #return(ds['audio_id'][exnum])
    return(exnum)
    

def recc(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)
        
    with torch.inference_mode():
        #wav = torch.from_numpy(wav).unsqueeze(0)
        #if torch.cuda.is_available():
        #    wav = wav.cuda()
        input_values = processor(wav,sampling_rate=16000).input_values[0]
        input_values = torch.tensor(input_values, device=device).unsqueeze(0)
        logits = model(input_values).logits
        pred_ids = torch.argmax(logits, dim=-1)
        xcp = processor.decode(pred_ids)
        return xcp
    

bl = gr.Blocks()
with bl:
    audio_file = gr.Audio(type="filepath")
    text_button = gr.Button("Recognise")
    text_output = gr.Textbox()
    
    text_button.click(recc, inputs=audio_file, outputs=text_output)


bl.launch()

#https://mercury-docs.readthedocs.io/en/latest/deploy/hugging-face-spaces/
#https://huggingface.co/spaces/pplonski/deploy-mercury
#https://discuss.huggingface.co/t/deploy-interactive-jupyter-notebook-on-spaces-with-mercury/17000
#https://huggingface.co/docs/transformers/notebooks