app.py

import gradio as gr
import torch
import sounddevice as sd
import numpy as np
from nemo.collections.asr.models import ASRModel

# Load the NeMo ASR model
model = ASRModel.from_pretrained("nvidia/canary-1b")
model.eval()

# Load the keyword spotting model
kws_model = torch.hub.load('snakers4/silero-vad', 'silero_vad')

# Constants
TRIGGER_WORD = "hey alexa"
TRIGGER_DURATION = 2  # Duration to record after trigger word is detected, in seconds
SAMPLE_RATE = 16000  # Sample rate for recording

def start_recording():
    print("Recording started...")
    audio = sd.rec(int(TRIGGER_DURATION * SAMPLE_RATE), samplerate=SAMPLE_RATE, channels=1, dtype='float32')
    sd.wait()
    return audio.flatten()

def detect_trigger(audio):
    # Perform keyword spotting
    is_triggered = kws_model(audio, sample_rate=SAMPLE_RATE) >= 0.5
    return is_triggered

def transcribe_triggered():
    while True:
        print("Listening for trigger word...")
        # Start recording
        recorded_audio = start_recording()

        # Check if trigger word is detected
        is_triggered = detect_trigger(recorded_audio)
        if is_triggered:
            print("Trigger word detected. Transcribing...")
            # Perform speech recognition
            transcription = model.transcribe([recorded_audio])
            return transcription[0]

iface = gr.Interface(transcribe_triggered, gr.inputs.NoInput(), "text", title="ASR with NeMo Canary Model (Triggered by 'Hey Alexa')")
iface.launch()