app.py

import gradio as gr
import numpy as np
import torch
import tempfile
import os
from scipy.io.wavfile import write
from transformers import (
    SpeechT5Processor,
    SpeechT5ForTextToSpeech,
    SpeechT5HifiGan
)

# =========================
# Model loading
# =========================
checkpoint = "Chithekitale/chichewa_tts_norules"

processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

# Make all keys consistent
speaker_embeddings = {
    "SPK1": "cmu_us_slt_arctic-wav-arctic_a0508.npy",
    "SPK2": "cmu_us_rms_arctic-wav-arctic_b0353.npy",
    "SPK3": "cmu_us_ksp_arctic-wav-arctic_b0087.npy",
    "SPK4": "cmu_us_rms_arctic-wav-arctic_b0353.npy",
    "SPK5": "cmu_us_slt_arctic-wav-arctic_a0508.npy",
}

SPEAKER_CHOICES = [
    "SPK1 (female)",
    "SPK2 (male)",
    "SPK3 (male)",
    "SPK4 (male)",
    "SPK5 (female)"
]

EXAMPLES = [
    ["Ndapita, koma ndibweranso pompano.", "SPK1 (female)"],
    ["Koma apapa zikuoneka kuti ziyenda bwino.", "SPK2 (male)"],
    ["Ineyo ndikuona kuti sizizasithanso.", "SPK3 (male)"],
    ["Mwina kusogolo kuno anthu ena azalimba mtima, koma panopana ndakaika.", "SPK4 (male)"],
    ["Simungasankhe munthu oti bola linamukana.", "SPK5 (female)"],
    ["Kodi chimanga panopa chikugulisidwa zingati, kapena nanunso simukudziwa?", "SPK5 (female)"],
]

SAMPLE_RATE = 16000

# =========================
# Helpers
# =========================
def get_speaker_key(speaker_label: str) -> str:
    # "SPK1 (female)" -> "SPK1"
    return speaker_label.split()[0]

def load_speaker_embedding(speaker: str) -> np.ndarray:
    speaker_key = get_speaker_key(speaker)

    if speaker_key not in speaker_embeddings:
        raise ValueError(f"Unknown speaker key: {speaker_key}")

    path = speaker_embeddings[speaker_key]

    try:
        speaker_embedding = np.load(path).astype(np.float32)
    except Exception as e:
        raise FileNotFoundError(
            f"Could not load speaker embedding file: {path}. Error: {e}"
        )

    if speaker_embedding.ndim == 2:
        speaker_embedding = speaker_embedding.mean(axis=0)

    speaker_embedding = np.squeeze(speaker_embedding)

    if speaker_embedding.shape != (512,):
        raise ValueError(
            f"Unexpected speaker embedding shape after processing: "
            f"{speaker_embedding.shape}. Expected (512,)"
        )

    return speaker_embedding

def save_audio_to_wav(audio: np.ndarray, sample_rate: int = SAMPLE_RATE) -> str:
    """
    Save generated int16 audio to a temporary WAV file and return its path.
    """
    temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
    temp_file.close()
    write(temp_file.name, sample_rate, audio)
    return temp_file.name

# =========================
# Inference
# =========================
def predict(text, speaker):
    try:
        if not text or len(text.strip()) == 0:
            return None, None, "Please enter some Chichewa text."

        inputs = processor(text=text, return_tensors="pt")
        input_ids = inputs["input_ids"][..., :model.config.max_text_positions]

        speaker_embedding = load_speaker_embedding(speaker)
        speaker_embedding = torch.tensor(
            speaker_embedding, dtype=torch.float32
        ).unsqueeze(0)

        with torch.no_grad():
            speech = model.generate_speech(
                input_ids,
                speaker_embedding,
                vocoder=vocoder
            )

        speech = speech.cpu().numpy()

        # Normalize safely before int16 conversion
        max_val = np.max(np.abs(speech))
        if max_val > 0:
            speech = speech / max_val

        speech = (speech * 32767).astype(np.int16)

        # Save WAV file for downloading
        wav_path = save_audio_to_wav(speech, SAMPLE_RATE)

        status = f"Generated speech successfully using speaker: {speaker}"
        return (SAMPLE_RATE, speech), wav_path, status

    except Exception as e:
        return None, None, f"Error during generation: {str(e)}"

def clear_all():
    return "", "SPK1 (female)", None, None, "Ready."

# =========================
# UI
# =========================
custom_css = """
.gradio-container {
    max-width: 1100px !important;
    margin: 0 auto;
}
.hero {
    text-align: center;
    padding: 10px 0 0 0;
}
.section-note {
    font-size: 0.95rem;
    opacity: 0.9;
}
"""

with gr.Blocks(css=custom_css, title="Baseline Chichewa Speech Synthesis Demo") as demo:
    gr.HTML(
        """
        <div class="hero">
            <h1>Baseline Chichewa Synthesis</h1>
            <p class="section-note">
                Enter Chichewa text, choose a speaker voice, and generate speech audio.
            </p>
        </div>
        """
    )

    with gr.Row():
        with gr.Column(scale=5):
            text_input = gr.Textbox(
                label="Input Text",
                placeholder="Type Chichewa text here...",
                lines=6
            )

            speaker_input = gr.Radio(
                label="Speaker Voice",
                choices=SPEAKER_CHOICES,
                value="SPK1 (female)"
            )

            with gr.Row():
                generate_btn = gr.Button("Generate Speech", variant="primary")
                clear_btn = gr.Button("Clear")

            status_box = gr.Textbox(
                label="System Status",
                value="Ready.",
                interactive=False
            )

        with gr.Column(scale=5):
            audio_output = gr.Audio(
                label="Generated Speech",
                type="numpy",
                autoplay=False
            )

            download_file = gr.File(
                label="Download Audio File"
            )

    gr.Markdown("### Example Inputs")
    gr.Examples(
        examples=EXAMPLES,
        inputs=[text_input, speaker_input]
    )

    generate_btn.click(
        fn=predict,
        inputs=[text_input, speaker_input],
        outputs=[audio_output, download_file, status_box],
        show_progress="full"
    )

    clear_btn.click(
        fn=clear_all,
        inputs=[],
        outputs=[text_input, speaker_input, audio_output, download_file, status_box]
    )

demo.launch()