Irodori TTS

The pitch is easy to state and hard to copy: put a 🔥 or 🎛 inside your Japanese sentence and the model changes how it is spoken. That removes the two usual ways to steer a TTS system, hand-written prosody tags and reference audio clips, and replaces them with something a writer can type in one keystroke. For scripted Japanese content, interactive fiction and character voices, it is a genuinely different authoring workflow. Strength one is range in one collection. You can start with Irodori-TTS-v4.1-Small (0.8B, updated Aug 11, 2026) and its Quantized and MF variants when your GPU is small, move up to Irodori-TTS-v4-Large (3B, updated 19 days ago) and its Quantized build when quality matters more than throughput, or drop back to Irodori-TTS-500M-v3 (0.5B, updated May 12, 2026) and Irodori-TTS-600M-v3-VoiceDesign (0.6B, updated May 31, 2026) when you want the older generation. The v2 pair (500M-v2 and 500M-v2-VoiceDesign, both updated Apr 11) is still listed for comparison. Strength two is that everything is downloadable and fine-tunable. Nothing here locks your custom speaker behind a hosted account: you fine-tune a checkpoint on your own recordings and own the result. Strength three is the audition path. Spaces demos exist for v4-Large, v4.1-Small, v3 and both v2 checkpoints, several running on Zero Agents, so you can hear a checkpoint before committing a download or a GPU hour. The 500M-v2 Demo page shows a configuration error in the current scrape, so treat that particular demo as unreliable. Weaknesses are structural, not quality-related. Irodori-TTS is Japanese-only, and it ships as model weights rather than a product, which means you own inference, latency, scaling and maintenance. Documentation in the collection is essentially the model listings themselves, so plan on reading model cards and experimenting rather than following a vendor guide. There is also no account, dashboard or API endpoint, so anything your team needs from a service provider, from uptime guarantees to a support channel, is outside this collection's scope. Where it fits: research on style-conditioned speech, prototypes with no speech budget, and any pipeline that already runs its own inference. Where it does not: multilingual products, non-technical teams that want a character roster in a finished app, and anything that requires an SLA.