Irodori-TTS is a Flow Matching-based text-to-speech model for Japanese, developed by the independent researcher Aratako and released under an MIT license covering both code and weights. Since February 2026 it has accumulated roughly 1,060 stars and 135 forks, and the codebase reached its v3 release this month. Its distinguishing feature is unusual enough to be worth stating up front: emoji placed in the input text influence how the speech is delivered. ## Emoji as a Prosody Channel Most expressive TTS systems control delivery through a separate style vector, a reference clip, or a natural-language style prompt. Irodori-TTS supports those too, but it also accepts emoji annotations inline in the text, and supported checkpoints use them to shape delivery and non-verbal vocal expressions — the laughs, sighs, and breaths that make synthetic speech sound less synthetic. The mechanism behind this is more interesting than the feature description suggests. The training corpus is an expanded Japanese speech dataset where the emoji annotations were generated automatically by a fine-tuned Qwen3-Omni-30B. In other words, a large multimodal model listened to the speech and labeled its emotional texture in emoji, and the TTS model learned to run that mapping in reverse. It is a pragmatic answer to the perennial shortage of emotion-labeled speech data. The practical appeal is that emoji are a control surface a user already knows how to operate. No style taxonomy to learn, no reference clip to record. ## Architecture The design largely follows Echo-TTS, generating into continuous DACVAE latents rather than discrete audio tokens. Five components make up the stack. A text encoder initialized from the pretrained llm-jp-3-150m, with self-attention, SwiGLU layers, and RoPE. A reference latent encoder that consumes patched reference audio latents for speaker identity. A caption encoder for style-control text covering emotion, tone, speaking style, and acoustic context. A Rectified Flow Diffusion Transformer built from joint-attention DiT blocks with Low-Rank AdaLN timestep conditioning, half-RoPE, and SwiGLU MLPs. And, new in v3, an integrated duration predictor that estimates output length automatically, removing the manual `--seconds` argument earlier versions required. Audio is represented through a 32-dimensional Semantic-DACVAE-Japanese codec that reconstructs 48 kHz waveforms — a high sample rate for an open TTS model of this size. ## Two Checkpoints, Three Modes The base model, Irodori-TTS-500M-v3, is about 500 million parameters and does zero-shot voice cloning from a reference clip, or unconditioned generation with `--no-ref`. The companion Irodori-TTS-600M-v3-VoiceDesign is a three-branch model that adds caption conditioning: give it a Japanese caption such as a request for a calm female voice, read softly at close distance, and it designs a voice to match without any reference audio. The third mode combines all of them — reference speech for identity plus a caption for style — which is the configuration that separates voice cloning from voice direction. Both checkpoints ship as safetensors on Hugging Face, each with a live Space demo. Generated audio is watermarked with Sony's SilentCipher when the library is available. The model card carries explicit ethical restrictions against impersonation, misinformation, and deepfakes. For a zero-shot voice cloning model, shipping watermarking on by default is the right posture and is still not the norm. ## Training and Tooling This is a full training repository rather than an inference drop. It supports distributed multi-GPU training via torchrun with gradient accumulation, bf16 mixed precision, and Weights & Biases logging; PEFT LoRA fine-tuning against the released checkpoints; and speaker inversion for learning reusable speaker embeddings. Backend coverage is broader than most TTS projects bother with. uv extras select CUDA 12.8, AMD ROCm, Intel XPU, or CPU/macOS MPS PyTorch builds, and the README documents a ROCm-specific workaround where `pytorch-triton-rocm` is required because `triton-rocm` alone does not provide `triton.language` for the transformers-to-dynamo import path. That is the kind of note that only appears after someone actually ran it on AMD hardware. A companion repository, Irodori-TTS-Server, provides an OpenAI-compatible inference API. Versioning is handled carefully: `main` tracks v3, `v2` and `v1` git tags preserve earlier codebases, v3 code remains backward-compatible with v2 checkpoints, and the README states plainly that v1 checkpoints and preprocessing are incompatible with v2 and v3. ## Pros and Cons The strengths start with licensing — MIT on both code and weights is about as permissive as this category gets, and many competing open TTS models restrict commercial use. Emoji-driven style control is a genuinely novel and low-friction interface. Full training, LoRA, and speaker-inversion code makes the project adaptable rather than merely usable. 48 kHz output via the DACVAE codec is high quality for a 500M model. Watermarking and stated ethical limits are present by default. Backend coverage across CUDA, ROCm, XPU, and CPU/MPS is unusually wide, and the automatic duration predictor removes a real usability wart from v2. The limitations are significant and should be weighed first. The model is Japanese-only — the text encoder is initialized from a Japanese LLM and the codec is Japanese-specific, so this is not a multilingual system with a Japanese mode. The model card publishes no quantitative evaluation: no MOS, no WER, no speaker-similarity numbers, so quality has to be judged by ear from the Spaces demos. The training dataset's sources and size are undisclosed, which matters for anyone assessing commercial or licensing risk despite the MIT tag. Emoji style control is qualified as affecting delivery in supported checkpoints rather than as a guaranteed deterministic control. Version churn from v1 to v3 in five months with a hard v1 compatibility break signals a fast-moving research codebase, and this is a single-maintainer project with 14 open issues — a bus factor worth acknowledging. ## Outlook Japanese TTS has historically lagged English in open models, and Irodori-TTS is one of the more technically current entries — Flow Matching over continuous latents is where the research frontier sits, several steps past the discrete-token autoregressive designs that dominated earlier open releases. The emoji-annotation trick is the more transferable idea: using a large audio-language model to generate emotional labels for unlabeled speech is a data-generation pattern that works for any language, and the approach is likely to be borrowed well beyond this repository. Whether the model itself grows depends largely on whether the maintainer keeps pace alone or attracts contributors. ## Conclusion Irodori-TTS is worth trying for Japanese voice synthesis work where expressive delivery matters and MIT licensing is a requirement — narration, character voices, or research on controllable prosody. The training and LoRA tooling makes it a reasonable base for fine-tuning on your own voices. Teams needing multilingual output, published quality benchmarks, or documented training-data provenance should treat it as a research project to watch rather than a production dependency, and should listen to the Hugging Face Space demos before investing setup time.