app.py

import gradio as gr
import torch
import librosa
import soundfile as sf
import tempfile
from unsloth import FastLanguageModel
import torch

from transformers import (
    AutoProcessor,
    AutoModelForImageTextToText,
    AutoTokenizer,
)

from unsloth import FastLanguageModel

# -----------------------------
# CONFIG
# -----------------------------
STT_MODEL_ID = "EpistemeAI/Audiogemma-3N-finetune"
TTS_MODEL_ID = "EpistemeAI/LexiVox"

TARGET_SR = 16000
MAX_TOKENS = 512

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.bfloat16 if DEVICE == "cuda" else torch.float32

# -----------------------------
# LOAD STT MODEL
# -----------------------------
print("Loading STT model...")
processor = AutoProcessor.from_pretrained(STT_MODEL_ID)

stt_model = AutoModelForImageTextToText.from_pretrained(
    STT_MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
)

stt_model.eval()

# -----------------------------
# LOAD TTS MODEL (UNSLOTH)
# -----------------------------
print("Loading TTS model with Unsloth...")

#tts_tokenizer = AutoTokenizer.from_pretrained(TTS_MODEL_ID)

tts_model, tts_tokenizer = FastLanguageModel.from_pretrained(
    model_name =TTS_MODEL_ID,
    max_seq_length= 2048, # Choose any for long context!
    dtype = None, # Select None for auto detection
    load_in_4bit = False, # Select True for 4bit which reduces memory usage
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

FastLanguageModel.for_inference(tts_model)
tts_model.eval()

# -----------------------------
# STT FUNCTION
# -----------------------------
def transcribe(audio_path):
    prompt = "Transcribe the audio accurately in German."

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "audio", "audio": audio_path},
                {"type": "text", "text": prompt},
            ],
        }
    ]

    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_tensors="pt",
        return_dict=True,
    )

    inputs = {k: v.to(stt_model.device) for k, v in inputs.items()}

    with torch.inference_mode():
        outputs = stt_model.generate(
            **inputs,
            max_new_tokens=MAX_TOKENS,
            do_sample=False,
            temperature=0.2,
        )

    text = processor.batch_decode(
        outputs,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )[0]

    return text

# -----------------------------
# SPEECH → SPEECH PIPELINE
# -----------------------------
def speech_to_speech(audio_file):
    if audio_file is None:
        return "", None

    # Ensure audio is valid
    _audio, _ = librosa.load(audio_file, sr=TARGET_SR)

    # ---------- STT ----------
    transcription = transcribe(audio_file)

    # ---------- TTS ----------
    tts_inputs = tts_tokenizer(
        transcription,
        return_tensors="pt",
    ).to(tts_model.device)

    with torch.inference_mode():
        speech_tokens = tts_model.generate(
            **tts_inputs,
            max_new_tokens=2048,
            do_sample=False,
            temperature=0.7,
        )

    audio_out = speech_tokens.cpu().numpy().squeeze()

    # Save temporary WAV
    tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
    sf.write(tmp.name, audio_out, TARGET_SR)

    return transcription, tmp.name

# -----------------------------
# GRADIO UI
# -----------------------------
with gr.Blocks(title="Audiogemma → LexiVox (Unsloth)") as demo:
    gr.Markdown(
        """
        # 🎙️ Speech → Text → Speech  
        **Audiogemma-3N + LexiVox (Unsloth Accelerated)**

        Upload audio or use your microphone.
        """
    )

    audio_input = gr.Audio(
        sources=["microphone", "upload"],
        type="filepath",
        label="Input Audio",
    )

    run_btn = gr.Button("Run Speech Loop")

    text_output = gr.Textbox(
        label="Transcription",
        lines=4,
    )

    audio_output = gr.Audio(
        label="Synthesized Speech",
        type="filepath",
    )

    run_btn.click(
        fn=speech_to_speech,
        inputs=audio_input,
        outputs=[text_output, audio_output],
    )

demo.launch()