app.py

# Swahili Text‑to‑Speech Gradio App – MP3 Output (mobile‑friendly)
# -----------------------------------------------------------------
# ‑ Generates clear Kiswahili speech and serves it **as an MP3 file** so that
#   iOS/Android browsers play it reliably.  Uses a fine‑tuned VITS checkpoint.
#
#‑ Dependencies (add to requirements.txt):
#   torch, transformers, gradio, scipy, pydub

import os
import tempfile
import torch
import numpy as np
import gradio as gr
import scipy.io.wavfile as wavfile
from pydub import AudioSegment
from transformers import VitsModel, AutoTokenizer

MODEL_NAME = "FarmerlineML/swahili-tts-2025"      # tokenizer
MODEL_CHECKPOINT = "FarmerlineML/Swahili-tts-2025_part4"  # acoustic model

device = "cuda" if torch.cuda.is_available() else "cpu"

# ---------- Load model ------------------------------------------------------
model = VitsModel.from_pretrained(MODEL_CHECKPOINT).to(device)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# Apply clear‑speech inference parameters once (no UI toggle)
model.noise_scale = 0.7
model.noise_scale_duration = 0.667
model.speaking_rate = 0.75  # must be >0 to avoid ZeroDivisionError

# ---------- Helper ----------------------------------------------------------

def _wav_to_mp3(wave_np: np.ndarray, sr: int) -> str:
    """Convert int16 numpy waveform to an MP3 temp file, return its path."""
    # Ensure int16 for pydub
    if wave_np.dtype != np.int16:
        # waveform from VITS is float32 in range [-1, 1]; scale and cast
        wave_np = (wave_np * 32767).astype(np.int16)

    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tf:
        wavfile.write(tf.name, sr, wave_np)
        wav_path = tf.name

    # Convert to mp3 via pydub (requires ffmpeg ‑ available in Spaces base img)
    mp3_path = wav_path.replace(".wav", ".mp3")
    AudioSegment.from_wav(wav_path).export(mp3_path, format="mp3", bitrate="64k")

    os.remove(wav_path)  # cleanup temp WAV
    return mp3_path


# ---------- TTS endpoint ----------------------------------------------------

def tts_generate(text: str):
    if not text:
        return None

    inputs = tokenizer(text, return_tensors="pt").to(device)
    with torch.no_grad():
        wave = model(**inputs).waveform[0].cpu().numpy()

    return _wav_to_mp3(wave, model.config.sampling_rate)

# ---------- UI --------------------------------------------------------------

examples = [
    ["zao kusaidia kuondoa umaskini na kujenga kampeni za mwamko wa virusi vya ukimwi amezitembelea"],
    ["Kidole hiki ni tofauti na vidole vingine kwa sababu mwelekeo wake ni wa pekee."],
    ["hivyo imekuwa msingi wa teknolojia yote ya umeme hasa nyaya za kila aina"],
    ["kumekuwa na majadiliano mengi juu ya usahihi wa ripoti hizi za madeni"],
    ["na kusaga ulipoanzia baada ya kumaliza masomo ndugu ruge mutahaba ndipo sasa mwishoni mwa"],
    ["Soko la Kariakoo huwa na watu wengi siku za Jumamosi."],
    ["Tafadhali hakikisha umefunga mlango kabla ya kuondoka."],
    ["Watoto walicheza mpira uwanjani hadi jua lilipotua."],
]

demo = gr.Interface(
    fn=tts_generate,
    inputs=gr.Textbox(lines=3, placeholder="Enter Swahili text here", label="Enter Swahili text here"),
    outputs=gr.Audio(type="filepath", label="Audio", autoplay=True),
    title="Swahili Text‑to‑Speech",
    description=(
        "Enter Swahili text and click **Submit** to play the audio"
    ),
    examples=examples,
    cache_examples=True,
)

if __name__ == "__main__":
    demo.launch()