Irodori-TTS-v4.1-Small GGUF for speech.cpp

These files run only in speech.cpp 0.8.0 or later. Their layout is speech.cpp's own, so llama.cpp, LM Studio, Ollama and other programs that read GGUF cannot run them.

It holds Aratako's Irodori-TTS v4.1-Small and its codec, Semantic-DACVAE-Japanese-32dim, in one file. It speaks Japanese, a sentence at a time, in the voice of a reference recording, in a voice described in words, or in a voice the model chooses, and passes the audio as the codec decodes it. It samples in Euler steps with the official runtime's guidance, 40 steps unless a request gives steps; 16 steps bring the first audio more than twice as soon. A request may also set the runtime's guidance, schedule and speaker scaling.

Use

Install speech.cpp (install), then:

# A voice file from a reference recording, then a sentence in that voice
speech voice irodori-tts me.wav me.voice.gguf
speech tts irodori-tts --add-voice me=me.voice.gguf --voice me --steps 16 -o out.wav "明日の東京は晴れです。"

# A voice described in words
speech tts irodori-tts --voice none --instructions "低く落ち着いた男性の声で、ゆっくりと読み上げてください。" --steps 16 \
    -o out.wav "明日の東京は晴れです。"

# A page on which to try it, and OpenAI's speech API
speech serve irodori-tts --open

irodori-tts is the name speech.cpp gives this model; speech fetches the F16 file from this repository the first time, and irodori-tts:q4_k or another type of the table below fetches that type instead. A path to the downloaded file works as well. The voices, the options, the worker and the HTTP API are in speech.cpp's documentation: Irodori-TTS, command line and server.

Files

In the Q6_K, Q5_K and Q4_K files the DiT's matrices whose rows are not whole blocks of 256 values are Q8_0, and two of the speaker encoder's, whose rows are not whole blocks of 32 either, are F16.

Voices

• A reference WAVE file of at most 120 s, encoded each time it is added.
• A voice file, which speech voice writes once from one or more references. It carries the codec's hash (67cb1a241c3b75d7…), so it works with this file and with v4.1-Small-MF's in any type.
• A voice file of a speaker-inversion embedding, which speech voice makes from the .speaker.safetensors file the official training saves, for this model alone.
• none: the model chooses the voice, or instructions describes it.

Accuracy

Every stage is checked against the official implementation's tensors, dumped on the CPU in float32, on an Apple M5:

On a GPU the audio is the same speech rather than the same waveform, since the GPU's matrix kernels round differently.

Each type, measured by speech-bench on 2026-10-12 on an Apple M5 with Metal, on the 20 sentences of speech-bench's prompts/speak-ja-JP.json spoken with seeds 1, 2 and 3 in a voice file of one reference, and heard by Qwen3-ASR 1.7B in Q8_0: the CER over the 60 takes. The memory is the peak memory footprint of speech for one request. The CER of one type moves by a few points from seed to seed, so only a larger difference tells two types apart.

Speed

The 20 sentences of speech-bench's prompts/speak-ja-JP.json through speech worker, in a voice file, one request at a time:

The first audio waits for the whole sentence's sampling, so it grows with the sentence.

Source

Converted by speech.cpp's reference/irodori-tts/convert.py rf, whose environment pins the official runtime at Aratako/Irodori-TTS@89f9d8f, from Aratako/Irodori-TTS-v4.1-Small at 2b28324dc263ed5e6638b3cf3dd94c82ead07b4b and Aratako/Semantic-DACVAE-Japanese-32dim at 47376ee24834d7a05a48ebabfe3cde29b3c5e214, to F32, from which speech quantize --type <type> wrote each file. The codec stays F32 in every type.

v4.1-Small-MF, which samples in 4 MeanFlow steps and gives the first audio sooner, is irodori-tts-mf in sakasegawa/Irodori-TTS-v4.1-Small-MF-GGUF.

License and ethical restrictions

Irodori-TTS v4.1-Small is Aratako's, released under the MIT License (LICENSE-MIT), with these ethical restrictions from its model card:

1. No Impersonation: Do not use this model to clone or impersonate the voice of any individual (e.g., voice actors, celebrities, public figures) without their explicit consent.
2. No Misinformation: Do not use this model to generate deepfakes or synthetic speech intended to mislead others or spread misinformation.
3. Voice Generation Disclaimer: When generating speech purely from text or captions without using reference audio, it is possible that the generated voice may coincidentally resemble that of a real person. This is strictly a probabilistic artifact within the latent space. The model was not trained with the intent of reproducing specific individuals.
4. Liability Disclaimer: The developers assume no liability for any misuse of this model. Users are solely responsible for ensuring their use of the generated content complies with applicable laws and regulations in their jurisdiction.

Semantic-DACVAE-Japanese-32dim is Aratako's, MIT on its model card. It is fine-tuned from Meta's facebook/dacvae-watermarked, which is under the Apache License 2.0 (LICENSE-APACHE-2.0). The SAM License that a sentence of that card's text names was corrected to Apache-2.0 in facebookresearch/dacvae on 2025-12-19.