adding-a-model.md

Adding a New Model

To add a new TTS model to the framework, create a new directory under models/ with three files.

1. Create the directory

mkdir models/mymodel
cd models/mymodel

2. manifest.json — Model metadata

{
    "name": "My Model",
    "voice_presets": ["voice1", "voice2"],
    "emotion_tags": ["<laugh>", "<sigh>"],
    "supports_description": false,
    "supports_reference_audio": true
}

Field	Type	Description
name	string	Display name shown in the TUI
voice_presets	string[]	Available voice names (empty if not supported)
emotion_tags	string[]	Tags users can embed in text (empty if not supported)
supports_description	bool	Whether the model accepts a voice description string
supports_reference_audio	bool	Whether the model can clone from a reference WAV

3. pyproject.toml — Dependencies

[project]
name = "tts-compare-mymodel"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
    "torch",
    "soundfile",
    "mymodel-package",
]

4. app.py — Model wrapper

The wrapper must:

1. Read JSON from stdin
2. Load the model and generate audio
3. Write a WAV file to the specified output path
4. Print timing data to stderr
5. Print the output file path to stdout

#!/usr/bin/env python3
import json
import sys
import time

import numpy as np
import soundfile as sf


def generate(text: str, output_path: str) -> str:
    t_load = time.perf_counter()
    print("Loading model...", file=sys.stderr)
    # model = load_your_model()
    load_ms = (time.perf_counter() - t_load) * 1000

    t_infer = time.perf_counter()
    print("Generating speech...", file=sys.stderr)
    # audio, sr = run_inference(model, text)
    infer_ms = (time.perf_counter() - t_infer) * 1000

    audio = np.clip(audio, -1.0, 1.0)
    audio = (audio * 32767).astype(np.int16)

    sf.write(output_path, audio, sr)
    print(
        f"TIMING:{json.dumps({'model_load_ms': round(load_ms, 1), 'model_inference_ms': round(infer_ms, 1)})}",
        file=sys.stderr,
    )
    print("Done.", file=sys.stderr)
    return output_path


if __name__ == "__main__":
    params = json.loads(sys.stdin.read())
    path = generate(
        text=params["text"],
        output_path=params["output_path"],
    )
    print(path)

5. Install and test

cd models/mymodel
uv sync                     # Creates isolated .venv

# Test directly
echo '{"text": "Hello", "output_path": "test.wav"}' | uv run python app.py

The model will automatically appear in the TUI on the next run of python main.py.