What is Higgs TTS 3?
Higgs TTS 3 is Boson AI's third-generation text-to-speech model, released on 4 June 2026 and listed on some leaderboards as Higgs Audio V3 TTS. It is a roughly 4-billion-parameter autoregressive decoder that predicts eight streams of audio codes at 25 frames a second and outputs 24 kHz speech. Boson reports a word or character error rate under 5% in 85 languages and between 5% and 10% in 17 more.
It clones a voice from a reference clip (a transcript of the clip improves the match) and accepts inline tags for 21 emotions, singing, shouting and whispering, speed, pitch, pauses and sound effects such as laughter or a sigh. The weights are free for research and non-commercial use; a Creator Use Grant also lets people monetise their own podcasts and videos with credit to Boson. Product use needs a commercial licence or Boson's API at $15 per million characters.
How much does Higgs TTS 3 cost?
| Plan | Price | What you get |
|---|---|---|
| Self-hosted weights | Free | Research, non-commercial and Creator Use Grant only |
| Boson API | $15 per million characters | Billed on input characters; $10 trial credit for new accounts |
| Commercial licence | Not published | Required to host, embed or resell the model |
Weights free for research, non-commercial use and credited creator channels; products need a commercial licence. Boson API: $15 per million characters, $10 trial credit.
Is Higgs TTS 3 worth it?
Higgs TTS 3 is the open model to look at for accurate speech in many languages, as long as you are not building a product on the weights. In Boson's tables it averages a 2.74% error rate across 32 languages on the MiniMax multilingual test, against 2.98% for OmniVoice and 5.15% for Fish Audio S2 Pro, and 3.61% across Boson's own 111-language set, where Qwen3-TTS-1.7B scores 97.09%. On EmergentTTS-Eval, a judged test of conversational delivery, it wins 53.65% of comparisons against 43.80% for Fish S2 Pro.
Those are the maker's numbers, and the listening arena tells a cooler story. Artificial Analysis lists Boson's Higgs Audio V3 TTS 78th of 104 at 1,039 Elo, below Kokoro's 1,064 and well behind Fish Audio Drama 3 at 1,275. Accuracy and preference are different things, and Higgs leads on the first. Boson's own breakdown also has it trailing OmniVoice on emotions (53.75% against 61.07%) and Qwen3-TTS on complex pronunciation.
It is built to serve. On one H100 with SGLang-Omni, Boson measures 617 ms mean latency and a real-time factor of 0.147 for a single request, rising to 1,079 ms at 16 concurrent requests while producing 61.84 seconds of audio per second. Streaming returns audio in chunks through an OpenAI-style endpoint. No VRAM figure is published.
The catch: the Research and Non-Commercial License forbids hosting, embedding in a product, production use and internal business use beyond evaluation, and it bars using outputs to train other speech models. It is also the strictest consent regime here: cloning a real person needs their explicit, verifiable consent, and high-risk commercial users must keep records Boson can ask to see. Nothing in the code enforces that, but the terms are clear.
Our take: Higgs TTS 3 suits a researcher or a podcaster crediting Boson; a company should treat the weights as a free trial of the $15-per-million-character API and compare that against Fish Audio at the same price.
Where it earns its place
A speech researcher at a university lab
Generate test audio in low-resource languages for an ASR benchmark
Run the weights on SGLang-Omni and render prompts across Boson's 102 single-digit-error languages. Research use is free; publishing results and samples in papers is allowed under the licence.
An independent podcaster producing a weekly show in English and Spanish
Voice ad reads and segments with laughter and pauses where the script calls for them
The Creator Use Grant covers monetised channels you own, provided the show notes credit Boson AI's Higgs Audio. Inline tags such as laughter or a long pause go straight into the script.
A customer-service platform evaluating voices for phone agents
Test multilingual agent speech before signing a contract
Internal evaluation is allowed on the weights; production needs the API at $15 per million characters, so 20 million characters a month would cost about $300.
Before you commit
- Setup
- Pull the SGLang-Omni Docker image, download the weights with a Hugging Face token and run sgl-omni serve; requests go to /v1/audio/speech. Supplying the reference clip's transcript improves cloning.
- Integrations
- SGLang-Omni with continuous batching, Hugging Face Transformers, Boson API with blocking and streaming generation, and Boson Workspace in the browser.
- Data and privacy
- Self-hosting keeps audio local. The licence prohibits cloning without explicit, verifiable consent and removing watermarks or provenance signals, and lets Boson ask distributors and hosts for an audit at most once a year.
- Cost at scale
- Boson's API bills $15 per million input characters, with generated audio not charged separately and style descriptions counted as input. Commercial self-hosting prices are not published.
- Limits
- 8,192-token training sequence length. 17 languages sit in the 5% to 10% error band, and latency rises from 617 ms to 1,079 ms as concurrency grows from 1 to 16 on an H100.
What’s good
- 2.74% average error across 32 languages on the MiniMax test, lowest in Boson's table
- Inline tags for 21 emotions, styles, pauses and sound effects
- 617 ms mean latency per request on one H100 with SGLang-Omni
- Clear consent rules, including record-keeping for high-risk uses
What’s not
- Weights cannot be used in products without a commercial licence
- 78th of 104 at 1,039 Elo on Artificial Analysis's arena
- Outputs may not be used to train other speech models
- No published VRAM requirement
Best for
Researchers and independent creators who want accurate, expressive speech in many languages, and companies trialling Boson's API before buying.
Higgs TTS 3 alternatives
Higgs TTS 3: common questions
- Is Higgs TTS 3 worth it?
- Higgs TTS 3 posts the lowest error rates in Boson's own multilingual tests, but listeners on Artificial Analysis rank it mid-table and its licence keeps it out of products. We rate it 3.8 out of 5.
- How much does Higgs TTS 3 cost?
- Weights free for research, non-commercial use and credited creator channels; products need a commercial licence. Boson API: $15 per million characters, $10 trial credit.
- What are the downsides of Higgs TTS 3?
- Weights cannot be used in products without a commercial licence; 78th of 104 at 1,039 Elo on Artificial Analysis's arena; outputs may not be used to train other speech models; no published VRAM requirement.
- Who is Higgs TTS 3 best for?
- Researchers and independent creators who want accurate, expressive speech in many languages, and companies trialling Boson's API before buying.
- How much does Higgs TTS 3 cost for a team or in production?
- Boson's API bills $15 per million input characters, with generated audio not charged separately and style descriptions counted as input. Commercial self-hosting prices are not published.
- Is Higgs TTS 3 safe for company data?
- Self-hosting keeps audio local. The licence prohibits cloning without explicit, verifiable consent and removing watermarks or provenance signals, and lets Boson ask distributors and hosts for an audit at most once a year.
- What are the best Higgs TTS 3 alternatives?
- The closest alternatives we have reviewed are Fish Audio Drama 3, OmniVoice, VoxCPM2 and Qwen3-TTS.
Answers checked Nov 7, 2026.