Engines¶

TTS-Studio ships with three interchangeable engines. All of them share the same CLI surface, chapter splitting, resumability, and output handling — pick one with --engine.

kokoro¶

The default engine: a small, high-quality local model that runs anywhere (Metal on Apple Silicon, CUDA, or CPU). It supports real-time streaming with --stream and per-chapter resumability.

1
2

uv sync --extra kokoro
uv run tts-studio convert book.epub --voice af_heart --speed 1.1

• --voice: Kokoro voice ID (default af_heart)
• --lang: a (en, en-us) or b (en-gb)

edge¶

A thin wrapper around edge-tts: fast, no model downloads, and a large catalogue of voices and languages. Handy when you don't want to pull in torch or MLX — it works with a bare uv sync.

1

uv run tts-studio convert input.txt --engine edge --voice en-US-AriaNeural

• --voice: Edge TTS voice name (default af_heart)
• Output is MP3 rather than WAV

breeze¶

Breeze TTS 2 on Apple Silicon, with voice design and cloning:

1
2
3
4
5
6
7
8
9

uv sync --extra breeze

# design a voice from a description
uv run tts-studio convert input.txt --engine breeze \
    --instruction "A warm, thoughtful young woman with a calm, reflective delivery" \
    --cfg-scale 4

# clone a voice from a 5-15s sample (transcript auto-transcribed with whisper)
uv run tts-studio convert input.txt --engine breeze --instruction sample.wav

Key options:

• --instruction: text voice description or a path to sample audio
• --ref-audio / --ref-text: reference audio plus its exact transcript
• --cfg-scale: CFG guidance scale for text instructions (try 4)
• --breeze-model: checkpoint dir or HF repo id (BREEZE_TTS_MODEL env var overrides)
• --seed: sampling seed (default 42)
• --breeze-workers: parallel chunk workers, each with its own model copy (~3 GB RAM)
• --breeze-depth-mode: cached or compiled depth-decoder mode

See Breeze TTS 2 notes on the home page for performance and voice-anchoring details.