Irodori-TTS-v4-Large: 3.29B Japanese TTS Guide & License

Irodori-TTS-v4-Large is a Japanese-only text-to-speech model of about 3.29B parameters, released on Hugging Face by Aratako. The repository was created on September 27, 2026. It unifies three conditions (input text, reference audio, and a caption) in one model, enabling zero-shot voice cloning, Voice Design (creating a voice from a text description), and voice cloning with style control. Emojis in the input text can also steer delivery and non-verbal sounds.

This article is based on the official model card, the quantized-variant model card, and the GitHub README as of October 1, 2026. Where the official sources are silent, such as required VRAM, we write 'not stated officially.' All benchmark figures are automatic evaluations by the developer; no large-scale human evaluation was conducted, so read the numbers with that in mind.

According to the model card, the model scales the v4 architecture from about 766M to about 3.29B parameters (roughly 4.3x). The method is a Rectified Flow Diffusion Transformer (RF-DiT), and the acknowledgments state that the design draws on Echo-TTS. Audio is represented with a codec using a 32-dimensional continuous latent and output at 48kHz. The basic specifications are as follows.

- Zero-shot voice cloning: pass a reference audio clip and the model reads the text in a voice close to that speaker
- Voice Design: create a voice from a text description (caption) alone, with no reference audio
- Voice cloning with style control: layer a caption on top of reference audio to specify emotion, delivery, and environment

In addition, emojis in the input text can control non-verbal sounds such as laughter, coughs, and sighs, as well as delivery (see below).

As described in the model card and README, v4-Large consists of five components. The input text and the caption are read by the same encoder, speaker characteristics are extracted from the reference audio, and the Diffusion Transformer generates the audio latent conditioned on all of these.

- Shared text/caption encoder: a fine-tuned T5Gemma 2 that reads both the input text and the caption
- Condition projectors: trained separately for text and for captions
- Reference latent encoder: conditions on speaker identity. Reference audio is up to 120 seconds in total
- Diffusion Transformer: a joint-attention DiT using Low-Rank AdaLN, half-RoPE, and SwiGLU MLPs
- Duration Predictor: automatically estimates the length of the output audio, built from SwiGLU MLP blocks

v4-Small used ModernBERT-ja (sbintuitions/modernbert-ja-310m) as its text encoder. v4-Large replaces it with one derived from the T5Gemma 2 text encoder, and this swap is also the reason for the license difference discussed below. The Duration Predictor is trained separately after the main model, with the other parameters frozen. The model card says this is the same approach as v4.1-Small.

The model card's 'What's New' lists four changes: scaling up, the T5Gemma 2 encoder, a separately trained Duration Predictor, and a higher Voice Design score. The table below sets confirmed values against v4.1-Small, the smaller model in the same Irodori-TTS series (a minor update of v4-Small).

Note that the Voice Design and voice cloning benchmarks compare v4-Large with v4-Small (not v4.1). The series history, by Hugging Face repository creation date, runs: 500M (2026-02-25), 500M-v2 (03-23), 500M-v2-VoiceDesign (03-30), 500M-v3 (05-12), 600M-v3-VoiceDesign (05-31), v4-Small (08-01), v4.1-Small (08-11), v4.1-Small-MF (09-12), and v4-Large (09-27).

According to the model card, the evaluation conditions were FP32 inference, 40 RF steps, and text CFG 3.0, reported as the mean and population standard deviation over 5 seeds (0 to 4). Joyo, JSUT, Coco-Nut, and JVS are not in the training data. Everything is an automatic evaluation by the developer, and the card states that no large-scale human MOS evaluation was run. First, Japanese reading (no reference audio, no caption).

The model card states that reading error rates are slightly higher for v4-Large than for v4.1-Small on both benchmarks. On reading accuracy alone, Small is ahead. Next is Voice Design (caption following). The 2,890 public voice descriptions in the Coco-Nut test set were combined with short, medium, and long texts, giving 8,670 clips per model (no reference audio), which Gemini 3.6 Flash scored from 1 to 5. This is a single run with seed 0.

The gain over v4-Small is +0.0652 (95% confidence interval [+0.0353, +0.0955]). The largest improvement was on long texts, from 3.9851 to 4.1343. The model card itself notes that this is a lightweight internal comparison using a single automatic judge, with no human verification, and not a research-grade benchmark. Voice cloning was evaluated on all 100 JVS speakers, 5 texts, 5 seeds, speaker CFG 5.0, broken down by reference audio length.

How to read this: cosine similarity is higher for v4-Large in all four conditions, but the top-1 at 120 seconds is slightly lower for v4-Large, and the gap is small. Most of the improvement is already obtained at about 30 seconds of reference audio, and similarity drops substantially with a single short clip. Where possible, use clean reference audio of about 30 seconds or more. To repeat, these are automatic evaluations with no human evaluation, and the cloning comparison is against v4-Small.

Neither the model card nor the quantized-variant model card states a required VRAM; this is not stated officially. All we have are checkpoint file sizes: the full-precision model.safetensors is about 13.15GB. Five post-training quantized variants made with torchao are published, listed below. File size is only a rough guide, since runtime usage adds reference audio, CFG branches, and decoding, so required VRAM cannot be derived from file size.

Execution of the quantized variants was verified on NVIDIA CUDA only. CPU, ROCm, and Intel XPU depend on PyTorch / torchao kernels and are untested in this release. Specify --model-precision bf16 at inference. To simply try the model first, the demo Space, which runs on ZeroGPU, is the easiest way to start in a browser.

The code is published on GitHub, and the README uses uv. Clone the repository, then sync with the extra that matches your backend.

The backend extras are mutually exclusive: cu128 (NVIDIA CUDA 12.8, Linux / Windows), rocm (AMD, Linux / WSL), xpu (Intel, Linux / Windows), and cpu (CPU only, or CPU / MPS on macOS). After syncing, run commands with uv run --no-sync. Below is the int8 inference command from the quantized-variant model card (no reference audio).

To use the full-precision model, replace --hf-checkpoint with Aratako/Irodori-TTS-v4-Large. The README examples are written with the v4.1-Small repository ID, so you need to substitute the ID when using v4-Large. The main options listed in the README are below (examples are written for v4.1-Small).

For browser-based operation there are Gradio UIs: gradio_app.py for voice cloning (port 7860) and gradio_app_voicedesign.py for Voice Design (7861). As of the README, both UIs default to v4.1-Small. As of October 1, 2026, the latest GitHub commit (2026-09-12) is titled 'Add MeanFlow distillation and v4-Large support,' and the configs include train_v4_large.yaml and train_v4_large_duration.yaml. On the other hand, v4-Large support for the OpenAI-compatible server (Irodori-TTS-Server), LoRA fine-tuning, and the MeanFlow distilled version is not stated officially; the README targets v4-Small / v4.1-Small for these.

Placing emojis in the input text controls delivery, emotion, and non-verbal sounds such as laughter, coughs, and sighs. EMOJI_ANNOTATIONS.md lists all 45; below is an excerpt of 12.

The README's example sentence is 「あははっ🤭、それ本当に言ってるの？…😮‍💨まぁ、君らしいけどね。」 with the caption 「余裕のある大人の男性。親しい相手に対して、くだけた雰囲気で呆れながらも楽しそうに話している。」 (a composed adult man talking casually to someone close, amused and a little exasperated). Repeating the same emoji strengthens the effect. The official docs state, however, that control is not perfect, and that the effect depends on context and can be inconsistent.

The biggest caveat for v4-Large is the license. According to the model card, because the shared encoder derives from google/t5gemma-2-1b-1b, the v4-Large weights fall under the Gemma Terms of Use, and use and redistribution must follow those terms and the Gemma Prohibited Use Policy. The repository bundles GEMMA_TERMS_OF_USE.md, GEMMA_PROHIBITED_USE_POLICY.md, and NOTICE. Even though the code is MIT, the weights are under a separate agreement, so the model cannot simply be called 'open source.' For the Gemma family itself, see our Gemma 4 introduction.

- Google claims no rights in Outputs
- On redistribution, you must include usage restrictions in your agreement, provide a copy of the terms, mark modifications, and include the NOTICE
- Google reserves the right to restrict, by remote or other means, use it reasonably believes violates the terms
- There is no clause banning commercial use, but it is conditional as above (the terms were last updated 2026-04-01)

- Do not clone or impersonate the voices of voice actors, celebrities, public figures, or others without their explicit consent
- Do not generate content for deepfakes or disinformation
- Even without reference audio, output may happen to resemble a real person (a probabilistic result; training was not intended to reproduce any specific individual)
- The developer takes no responsibility for misuse, and legal compliance is the user's responsibility

None of this is legal advice. If you plan commercial use or redistribution, read the original Gemma Terms of Use and Prohibited Use Policy before making a final decision. Voice cloning presupposes the consent of the person whose voice is used as the reference.

v4-Large trails Small on reading accuracy, so bigger is not always better. It is the model to pick when voice specification and reproduction matter. In positioning, it is close to local voice tools such as VoiceStudio in that it can run locally. For a cloud TTS API example, there is the xAI Grok voice API; check each provider's official documentation for pricing, language support, and latency. For the reverse direction, fully local speech input (STT), see TypeWhisper.

- Have a person listen for misreadings. The official docs state that rare proper nouns, technical terms, and context-dependent readings can be misread
- Get the speaker's consent and use clean reference audio of about 30 seconds or more. With a single short clip, speaker similarity drops substantially
- Avoid contradictions between the caption and the reference audio. Contradictions can cause instability or unnatural output, or one side may take priority. Use the caption for emotion, style, and environment, and match voice quality to the reference audio
- Do not strip the watermark. Disclose that the audio is synthetic when you publish it
- Check the license. The v4-Large weights are under the Gemma Terms of Use, which carry obligations on redistribution
- Japanese only. It cannot be used for narration in other languages

For business uses such as phone answering, misreadings and disclosure of synthetic audio are especially important. For a cost overview, see AI phone answering cost guide.

- Aratako/Irodori-TTS-v4-Large (model card)
- Aratako/Irodori-TTS-v4-Large-Quantized (quantized variants)
- Irodori-TTS-v4-Large-Demo (demo Space)
- GitHub: Aratako/Irodori-TTS
- EMOJI_ANNOTATIONS.md (emoji list)
- Aratako/Irodori-TTS-v4.1-Small
- Gemma Terms of Use
- GitHub: Aratako/Irodori-TTS-Server (OpenAI-compatible server)