app.py

import nltk
nltk.download("punkt")

import random
import numpy as np
import torch
import io
import os
import soundfile as sf
from nltk.tokenize import sent_tokenize
from pydub import AudioSegment, silence  # Added silence module
import gradio as gr

from chatterbox.src.chatterbox.tts import ChatterboxTTS

# ===============================
# DEVICE
# ===============================
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Running on: {DEVICE}")

# ===============================
# LOAD MODEL ONCE
# ===============================
MODEL = None

def get_model():
    global MODEL
    if MODEL is None:
        print("Loading Chatterbox model...")
        MODEL = ChatterboxTTS.from_pretrained(DEVICE)
        if hasattr(MODEL, "to"):
            MODEL.to(DEVICE)
        print("Model ready.")
    return MODEL

get_model()

# ===============================
# SEED
# ===============================
def set_seed(seed):
    torch.manual_seed(seed)
    if DEVICE == "cuda":
        torch.cuda.manual_seed_all(seed)
    random.seed(seed)
    np.random.seed(seed)

# ===============================
# PODCAST SAFE SETTINGS
# ===============================
MAX_CHARS = 220
SILENCE_MS = 250   # Reduced slightly since we are cleaning audio
FADE_IN = 10       # Reduced fade to avoid eating words
FADE_OUT = 10      # Reduced fade to avoid weird half-breath sounds

# ===============================
# HELPER: TRIM SILENCE/BREATHS
# ===============================
def trim_audio_segment(audio_segment, silence_thresh=-40):
    """
    Trims silence or quiet breath sounds from the start and end of a chunk.
    Adjust silence_thresh (dBFS) if it cuts off actual words.
    """
    # Detect non-silent chunks
    non_silent_ranges = silence.detect_nonsilent(
        audio_segment, 
        min_silence_len=100, 
        silence_thresh=silence_thresh
    )

    # If audio is completely silent or empty, return empty
    if not non_silent_ranges:
        return AudioSegment.empty()

    # Get start of first sound and end of last sound
    start_trim = non_silent_ranges[0][0]
    end_trim = non_silent_ranges[-1][1]

    return audio_segment[start_trim:end_trim]

# ===============================
# MAIN TTS FUNCTION
# ===============================
def generate_tts(
    text,
    ref_audio=None,
    exaggeration=0.4,
    temperature=0.7,
    seed=0,
    cfg_weight=0.6,
):

    model = get_model()

    if seed != 0:
        set_seed(int(seed))

    kwargs = {
        "exaggeration": exaggeration,
        "temperature": temperature,
        "cfg_weight": cfg_weight,
    }

    # --------------------------------
    # Handle reference voice
    # --------------------------------
    temp_prompt = None
    if ref_audio:
        try:
            audio = AudioSegment.from_file(ref_audio)
            temp_prompt = "voice_prompt.wav"
            audio.export(temp_prompt, format="wav")
            kwargs["audio_prompt_path"] = temp_prompt
        except:
            print("Reference audio failed — using default voice.")

    # --------------------------------
    # Sentence chunking
    # --------------------------------
    sentences = sent_tokenize(text)

    chunks = []
    current = ""

    for s in sentences:
        if len(current) + len(s) < MAX_CHARS:
            current += " " + s
        else:
            chunks.append(current.strip())
            current = s

    if current.strip():
        chunks.append(current.strip())

    print(f"Total chunks: {len(chunks)}")

    # --------------------------------
    # Generate audio per chunk
    # --------------------------------
    final_audio = AudioSegment.empty()
    clean_pause = AudioSegment.silent(duration=SILENCE_MS)

    for i, chunk in enumerate(chunks):
        print(f"Generating chunk {i+1}/{len(chunks)}")

        # 1. Generate Raw Audio
        wav = model.generate(chunk, **kwargs)
        wav_np = wav.squeeze(0).cpu().numpy()

        buffer = io.BytesIO()
        sf.write(buffer, wav_np, model.sr, format="WAV")
        buffer.seek(0)

        segment = AudioSegment.from_wav(buffer)

        # 2. TRIM ARTIFACTS (The Fix)
        # We strip the "trailing breath" or silence from the model output
        # BEFORE we add our own clean silence.
        segment = trim_audio_segment(segment, silence_thresh=-45)

        # 3. Apply light fade only after trimming
        if len(segment) > 0:
            segment = segment.fade_in(FADE_IN).fade_out(FADE_OUT)
            final_audio += segment + clean_pause

    # --------------------------------
    # Export
    # --------------------------------
    output_path = "story_voice.mp3"
    final_audio.export(output_path, format="mp3", bitrate="192k")

    if temp_prompt and os.path.exists(temp_prompt):
        os.remove(temp_prompt)

    return output_path

# ===============================
# GRADIO UI
# ===============================
with gr.Blocks() as demo:
    gr.Markdown("## 🎙️ Storyteller / Podcast Chatterbox TTS (Cleaned)")

    text = gr.Textbox(
        label="Story Text",
        lines=12,
        placeholder="Paste your full story here..."
    )

    ref = gr.Audio(
        sources=["upload", "microphone"],
        type="filepath",
        label="Reference Voice (optional)"
    )

    exaggeration = gr.Slider(0.25, 1.0, value=0.4, step=0.05, label="Emotion")
    temperature = gr.Slider(0.3, 1.2, value=0.7, step=0.05, label="Variation")
    cfg = gr.Slider(0.3, 1.0, value=0.6, step=0.05, label="Voice Stability")

    seed = gr.Number(value=0, label="Seed (0 = random)")

    btn = gr.Button("Generate Voice")
    out = gr.Audio(label="Final Audio")

    btn.click(
        fn=generate_tts,
        inputs=[text, ref, exaggeration, temperature, seed, cfg],
        outputs=out
    )

demo.launch(share=True)