app.py

import streamlit as st
import torch
import base64
import tempfile
import os
from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor

# Setup model
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model_id = "KBLab/kb-whisper-tiny"

@st.cache_resource
def load_model():
    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        model_id, torch_dtype=torch_dtype, use_safetensors=True, cache_dir="cache"
    )
    model.to(device)
    processor = AutoProcessor.from_pretrained(model_id)
    return pipeline(
        "automatic-speech-recognition",
        model=model,
        tokenizer=processor.tokenizer,
        feature_extractor=processor.feature_extractor,
        torch_dtype=torch_dtype,
        device=device,
    )

asr_pipeline = load_model()

st.title("Swedish Speech-to-Text Demo")

# Audio Upload Option
uploaded_file = st.file_uploader("Ladda upp en ljudfil", type=["wav", "mp3", "flac"])

# JavaScript for recording audio
audio_recorder_js = """
<script>
let mediaRecorder;
let audioChunks = [];
let isRecording = false;

function startRecording() {
    if (!isRecording) {
        isRecording = true;
        navigator.mediaDevices.getUserMedia({ audio: true }).then(stream => {
            mediaRecorder = new MediaRecorder(stream);
            audioChunks = [];
            mediaRecorder.ondataavailable = event => {
                audioChunks.push(event.data);
            };
            mediaRecorder.onstop = () => {
                const audioBlob = new Blob(audioChunks, { type: 'audio/wav' });
                const reader = new FileReader();
                reader.readAsDataURL(audioBlob);
                reader.onloadend = () => {
                    const base64Audio = reader.result.split(',')[1];
                    fetch('/save_audio', {
                        method: 'POST',
                        headers: { 'Content-Type': 'application/json' },
                        body: JSON.stringify({ audio: base64Audio })
                    }).then(response => response.json()).then(data => {
                        console.log(data);
                        window.location.reload();
                    });
                };
            };
            mediaRecorder.start();
        });
    }
}

function stopRecording() {
    if (isRecording) {
        isRecording = false;
        mediaRecorder.stop();
    }
}
</script>

<button onclick="startRecording()">🎤 Starta inspelning</button>
<button onclick="stopRecording()">⏹️ Stoppa inspelning</button>
"""

st.components.v1.html(audio_recorder_js)

# Processing audio input (uploaded file or recorded)
audio_path = None

if uploaded_file is not None:
    # Save uploaded file to a temp location
    with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(uploaded_file.name)[-1]) as temp_audio:
        temp_audio.write(uploaded_file.read())
        audio_path = temp_audio.name

elif "audio_data" in st.session_state and st.session_state["audio_data"]:
    # Decode base64 audio from JavaScript recording
    audio_bytes = base64.b64decode(st.session_state["audio_data"])
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio:
        temp_audio.write(audio_bytes)
        audio_path = temp_audio.name

# Transcribe if we have audio
if audio_path:
    st.audio(audio_path, format="audio/wav")
    
    with st.spinner("Transkriberar..."):
        transcription = asr_pipeline(audio_path)["text"]

    st.subheader("📜 Transkription:")
    st.write(transcription)

    # Cleanup temp file
    os.remove(audio_path)