maneko 真似猫

This project exists purely because I am too cheap to upgrade my (still working perfectly fine) Intel Macbook Pro to Apple Silicon. It's a native Rust/Candle TTS engine that can run pocket-tts v2 and irodori-tts v3 on those really dated Intel mac, slowly.

And it turns out that only the pocket-tts is worth running and can get sub-1x real time audio output, while irodori-tts is 5~10x slower than real time. It's still a win and a good way to waste tokens.

Usage

Weights load from a HuggingFace cache — point HF_HOME at the cache holding that engine's repos (both engines use the project-local .cache/huggingface). Always build --release (debug Candle is ~40× slower); --features accelerate on Apple Silicon CPU, --features metal for the GPU.

CLI (tts):

# pocket-tts (multilingual)
HF_HOME=$PWD/.cache/huggingface \
  cargo run --release --features accelerate -p tts-cli -- \
  generate --engine pocket --language german --voice voices/de/nathan.wav -o de.wav --text "Hallo Welt."

# Irodori v3 (Japanese, voice cloning) — duration is auto-predicted; omit --seconds, or pass it to override
HF_HOME=$PWD/.cache/huggingface \
  cargo run --release --features accelerate -p tts-cli -- \
  generate --engine irodori --voice voices/ja/ref.wav --steps 8 -o ja.wav --text "こんにちは。"

Python (maneko): build with maturin develop --features accelerate (in crates/interfaces/python), then:

import maneko
p = maneko.Pocket()
maneko.save_wav("out.wav", p.generate("Hello.", language="english_2026-04", voice="alba"), p.sample_rate("english_2026-04"))

i = maneko.Irodori()
maneko.save_wav("ja.wav", i.generate("こんにちは。", voice="ref.wav", seconds=4, steps=8), i.sample_rate)

Rust: depend on pocket / irodori; the frozen entry points are pocket::Engine (generate(text, language, voice)) and irodori::Irodori (generate(text, ref_wav, opts)).

License

AGPL-3.0-or-later (see LICENSE). maneko incorporates and ports MIT-licensed upstreams (babybirdprd/pocket-tts, mlx-audio); their notices are retained in NOTICE (MIT/Apache are AGPL-compatible). Model weights are separate artifacts under their own terms (Kyutai/Aratako/llm-jp), including no non-consensual voice cloning.