Pranaym12/2-LiveASR

1import gradio as gr2import torch3import time4import librosa5import soundfile6import nemo.collections.asr as nemo_asr7import tempfile8import os9import uuid10 11from transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration12import torch13 14# PersistDataset -----15import os16import csv17import gradio as gr18from gradio import inputs, outputs19import huggingface_hub20from huggingface_hub import Repository, hf_hub_download, upload_file21from datetime import datetime22 23# ---------------------------------------------24# Dataset and Token links - change awacke1 to your own HF id, and add a HF_TOKEN copy to your repo for write permissions25# This should allow you to save your results to your own Dataset hosted on HF. 26 27DATASET_REPO_URL = "https://huggingface.co/datasets/awacke1/ASRLive.csv"28DATASET_REPO_ID = "awacke1/ASRLive.csv"29DATA_FILENAME = "ASRLive.csv"30DATA_FILE = os.path.join("data", DATA_FILENAME)31HF_TOKEN = os.environ.get("HF_TOKEN")32 33PersistToDataset = False34#PersistToDataset = True  # uncomment to save inference output to ASRLive.csv dataset35 36if PersistToDataset:37    try:38        hf_hub_download(39            repo_id=DATASET_REPO_ID,40            filename=DATA_FILENAME,41            cache_dir=DATA_DIRNAME,42            force_filename=DATA_FILENAME43        )44    except:45        print("file not found")46    repo = Repository(47        local_dir="data", clone_from=DATASET_REPO_URL, use_auth_token=HF_TOKEN48    )49           50def store_message(name: str, message: str):51    if name and message:52        with open(DATA_FILE, "a") as csvfile:53            writer = csv.DictWriter(csvfile, fieldnames=["name", "message", "time"])54            writer.writerow(55                {"name": name.strip(), "message": message.strip(), "time": str(datetime.now())}56            )57        # uncomment line below to begin saving - 58        commit_url = repo.push_to_hub()59        ret = ""60        with open(DATA_FILE, "r") as csvfile:61            reader = csv.DictReader(csvfile)62            63            for row in reader:64                ret += row65                ret += "\r\n"66    return ret            67 68# main -------------------------69mname = "facebook/blenderbot-400M-distill"70model = BlenderbotForConditionalGeneration.from_pretrained(mname)71tokenizer = BlenderbotTokenizer.from_pretrained(mname)72 73def take_last_tokens(inputs, note_history, history):74    filterTokenCount = 128 # filter last 128 tokens75    if inputs['input_ids'].shape[1] > filterTokenCount:76        inputs['input_ids'] = torch.tensor([inputs['input_ids'][0][-filterTokenCount:].tolist()])77        inputs['attention_mask'] = torch.tensor([inputs['attention_mask'][0][-filterTokenCount:].tolist()])78        note_history = ['</s> <s>'.join(note_history[0].split('</s> <s>')[2:])]79        history = history[1:]80    return inputs, note_history, history81 82def add_note_to_history(note, note_history):83    note_history.append(note)84    note_history = '</s> <s>'.join(note_history)85    return [note_history]86 87 88 89SAMPLE_RATE = 1600090model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained("nvidia/stt_en_conformer_transducer_xlarge")91model.change_decoding_strategy(None)92model.eval()93 94def process_audio_file(file):95    data, sr = librosa.load(file)96    if sr != SAMPLE_RATE:97        data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)98    data = librosa.to_mono(data)99    return data100 101 102def transcribe(audio, state = ""):   103    if state is None:104        state = ""105    audio_data = process_audio_file(audio)106    with tempfile.TemporaryDirectory() as tmpdir:107        audio_path = os.path.join(tmpdir, f'audio_{uuid.uuid4()}.wav')108        soundfile.write(audio_path, audio_data, SAMPLE_RATE)109        transcriptions = model.transcribe([audio_path])110        if type(transcriptions) == tuple and len(transcriptions) == 2:111            transcriptions = transcriptions[0]112        transcriptions = transcriptions[0]113        114    if PersistToDataset:115        ret = store_message(transcriptions, state) # Save to dataset - uncomment to store into a dataset - hint you will need your HF_TOKEN116        state = state + transcriptions + " " + ret117    else:118        state = state + transcriptions119    return state, state120 121gr.Interface(122    fn=transcribe,123    inputs=[124        gr.Audio(source="microphone", type='filepath', streaming=True),125        "state",126    ],127    outputs=[128        "textbox",129        "state"130    ],131    layout="horizontal",132    theme="huggingface",133    title="🗣️ASR-Gradio-Live🧠💾",134    description=f"Live Automatic Speech Recognition (ASR).",135    allow_flagging='never',136    live=True,    137    article=f"Result💾 Dataset: [{DATASET_REPO_URL}]({DATASET_REPO_URL})"138).launch(debug=True)139