maneko

maneko — quantized TTS weights

Ready-to-run quantized weights for maneko, a Rust/Candle TTS engine, so you don't have to quantize locally. Two TTS model families, each self-contained, plus a speech-restoration preprocessor.

pocket-tts/ — pocket-tts v2, Q8_0

Q8_0 quantizations of the pocket-tts v2 per-language models (a derivative of kyutai/pocket-tts). One GGUF per language:

pocket-tts/languages/<lang>/model.q8.gguf
  english  english_2026-01  english_2026-04
  german   german_24l   french_24l   italian  italian_24l
  portuguese  portuguese_24l  spanish  spanish_24l

• What's quantized: Linear weights → Q8_0; conv / norm / embedding / bias kept F16 (the source dtype, so dequant→f32 is numerically identical). ~1.6–1.8× smaller than the f16 source (6-layer ~121–137 MB, 24-layer _24l ~351 MB).
• Quality: Q8_0 is transparent — word-for-word identical to f32 on Whisper round-trips where decoding is deterministic, and indistinguishable native speech at normal temperature.
• Voices: the upstream per-language built-in voices work (the engine imports their precomputed model_state). Native voices: juergen (de), lola (es), estelle (fr), giovanni (it), rafael (pt); the rest are English.

Usage (maneko)

let model = pocket::TTSModel::load_gguf("german_24l", "pocket-tts/languages/german_24l/model.q8.gguf")?;

irodori-tts/ — Irodori v3 (Japanese, 48 kHz)

A self-contained bundle for the Irodori engine — q8 DiT + f16 codec + tokenizer, so maneko needs no third-party repo at runtime:

irodori-tts/
  model.q8.gguf            # v3 DiT — Q8_0 Linear weights (F16 elsewhere), ~610 MB
  dacvae.f16.safetensors   # Semantic-DACVAE codec, f16, ~215 MB
  tokenizer.json           # llm-jp tokenizer (unmodified)

• DiT — Q8_0 quantization of the Irodori v3 DiT (derivative of Aratako/Irodori-TTS-500M-v3), ~4× smaller than f32. v3's integrated duration predictor auto-lengths each clip (no manual seconds).
• DACVAE — f16 of the Semantic-DACVAE-Japanese codec (Aratako/Semantic-DACVAE-Japanese-32dim), half the f32 size. maneko runs it f16 on Metal, f32 on CPU.
• Quality: q8 DiT + f16 codec are transparent on a Whisper round-trip vs the f32 originals.

Usage (maneko)

let iro = irodori::Irodori::from_hf(&device)?; // pulls this irodori-tts/ bundle

sidon/ — Sidon speech restoration (not TTS)

Not a TTS engine — it synthesises nothing from text. Sidon (sarulab-speech, U-Tokyo) turns noisy in-the-wild speech into studio-quality 48 kHz audio, in dozens of languages. In maneko its job is cleaning the reference wav that the voice-cloning engines imitate — clone quality is capped by reference quality — and standalone offline restoration.

sidon/sidon.safetensors    # both stages in one f32 file, ~984 MB

One file, 407 tensors, 246 053 490 params:

• feature predictor (193.6 M) — facebook/w2v-bert-2.0 truncated to its first 8 conformer layers, with Sidon's LoRA (r=64, α=16) merged in, so no adapter is needed at runtime. Cleanses SSL features at 50 Hz.
• vocoder (52.4 M) — a DAC-style Snake conv decoder, Decoder(input_channel=1024, channels=1536, rates=[8,5,4,3,2]), 960× upsample to 48 kHz. Weight-norm is kept split as weight_g/weight_v and folded at load.

Why this file rather than the upstream repos. Upstream ships the deployed model as TorchScript (sarulab-speech/sidon-v0.1) and the weights as a PEFT adapter (sarulab-speech/sidon_raw_weight). The adapter was saved with bias="lora_only", which marks the target modules' biases trainable but does not persist them — base + adapter reproduces the deployed encoder only to ~2.6e-2, and nothing warns you. This file adds the 16 trained output_dense.bias vectors recovered from the TorchScript module, and reproduces the deployed feature predictor to 1.07e-06.

Fidelity. Against the official demo Space's own output on an 11 s clip: correlation 0.999916, SNR 37.7 dB — the same score the PyTorch reference itself gets against that demo.

Speed. ~3× slower than real time on an Intel i9-9880H CPU (~85 % of it the vocoder, which is memory-bandwidth bound). Fine for one-shot reference cleanup, which is cached; not a realtime path.

Usage (maneko)

let sidon = sidon::Sidon::load("sidon/sidon.safetensors", &device)?;
let clean = sidon.restore(&noisy, sample_rate, &device)?; // -> 48 kHz

License & attribution

Derivatives of permissively-licensed upstream models, redistributed with attribution:

• pocket-tts/ — quantized derivative of kyutai/pocket-tts, CC-BY-4.0 (released under the same).
• irodori-tts/ — derivatives of Aratako/Irodori-TTS-500M-v3 and Aratako/Semantic-DACVAE-Japanese-32dim (MIT), plus the unmodified llm-jp/llm-jp-3-150m tokenizer (Apache-2.0).
• sidon/ — derivative of sarulab-speech/sidon-v0.1 and sarulab-speech/sidon_raw_weight, fine-tuned from facebook/w2v-bert-2.0 — all MIT, redistributed under the same. Cite the paper: Sidon: Fast and Robust Open-Source Multilingual Speech Restoration for Large-scale Dataset Cleansing (arXiv:2509.17052).

Each GGUF also carries general.{license,source.url,description} metadata.

Acceptable use: the upstream policies carry over — do not use these models for non-consensual voice cloning or impersonation.