Boson AI drops Higgs TTS 3 with inline actor controls

The most frustrating part of using AI voices has always been the complete lack of stage presence. Traditional models are glorified readers: they plow through text without knowing when to drop their tone, hold a dramatic pause, or let a sigh slip through.

Boson AI launched Higgs TTS 3, a new speech model tailored specifically for real-time conversational agents, and posted the weights publicly on Hugging Face. The standout trick is that it lets you direct the voice mid-sentence directly through inline tags.

Directing an AI actor mid-sentence

In most speech pipelines, text-to-speech is treated as an afterthought tacked onto the end of generation. Higgs TTS 3 approaches speech as part of the live conversation itself.

Developers can insert control tags directly into the text stream. Without interrupting generation, you can tweak emotions, delivery style, speed, pitch, micro-pauses, and sound effects right in the middle of an utterance.

If your agent gets hit with a surprising question, it can hesitate, adjust its tone, and deliver an answer that actually sounds conversational. For developers producing audio dramas or games, it feels less like tweaking parameters and more like directing a voice actor in real time.

The model supports both streaming and blocking modes, handles reference-based voice cloning, and plugs cleanly into SGLang-Omni for fast local serving.

Dialing in precision across 100+ languages

Expressiveness is cheap if the voice bot constantly mangles foreign phrases or complex syntax, so Boson ran the model through extensive benchmarks.

Higgs TTS 3 reached single-digit WER and CER across more than 100 languages. In evaluations across suites like Seed-TTS, CV3, MiniMax-Multilingual, and Boson's internal 111-language benchmark, it posted lower error rates than Higgs TTS 2.

The benchmark comparisons spanned notable open and commercial contenders, including Fish Audio S2 Pro, Qwen3-TTS-1.7B, VibeVoice-7B, and IndexTTS-2. Boson also ran behavioral tests for conversational cues like questions, syntactic complexity, and paralinguistic sounds, with results reproduced independently by the SGLang-Omni team.

Anyone curious can test the voice sliders in the Boson Workspace browser demo, call it via API, or pull the model directly from Hugging Face for self-hosting.

Why it matters

Voice AI has pretty much solved pristine audio fidelity, but it has been consistently terrible at human nuance.

For builders creating interactive agents, audio dramas, or NPC dialogue, Higgs TTS 3 solves a massive headache. You no longer have to play prompt roulette or splice together ten separate audio takes just to get a character to whisper a punchline. Writing the inflection into the text script cuts production time dramatically.

Releasing the model weights on Hugging Face is also a loud shot across the bow of closed speech APIs. High-control emotional synthesis used to be an expensive proprietary moat. By handing capable weights to the open-source community, Boson puts immediate pressure on commercial voice vendors to offer deeper controls at lower prices.

We think hooking dynamic controls directly into the LLM's text output stream is brilliant because it avoids building an awkward secondary control layer. The main risk is formatting fragility: if an upstream LLM hallucinates an invalid tag mid-stream, the speech engine might stumble. But as a foundation for real-time voice agents, it makes AI speech feel dramatically more alive.