Higgs TTS 3

Boson AI's 4.65B text-to-speech model for voice chat: 100+ languages, cloning, inline emotion and sound-effect tags, non-commercial license.

In short. Higgs TTS 3 turns text into expressive speech in more than 100 languages and can copy a voice from a short recording. You can put tags in the text for emotions, whispering, laughter, speed and pauses. The weights are free for research and personal use, and for creators who credit Boson, but a product or hosted service needs a paid license.

Summary

Higgs TTS 3 is Boson AI's text-to-speech model for voice chat. It first appeared on Hugging Face on 4 June 2026 as Higgs Audio v3 and took the name Higgs TTS 3 on 9 July 2026, when its license gained a creator grant. It is a 4.65B-parameter autoregressive decoder that reads interleaved text and audio tokens, with audio coded as 8 codebooks at 25 frames per second and 24 kHz output, and a training sequence length of 8,192 tokens. Boson says it reaches single-digit word or character error rates in 102 languages, 85 of them under 5, and it clones a voice zero-shot from a reference clip. Tags written into the text set 21 emotions, singing, shouting or whispering, nine sound effects such as laughter, coughs and sighs, and speed, pitch and pauses.

On Boson's own macro-averaged word or character error rates, where lower is better, it scores 1.11 on Seed-TTS (OmniVoice 1.21, Qwen3-TTS-1.7B 1.30, Fish Audio S2 Pro 1.31), 4.41 on CV3 (Fish Audio S2 Pro 4.60), 2.74 on MiniMax-Multilingual (OmniVoice 2.98) and 3.61 on Boson's internal 111-language set (OmniVoice 3.63, Fish Audio S2 Pro 8.68). On Emergent TTS it wins 53.65% of comparisons against the benchmark's baseline, ahead of MOSS-TTS-v1.5 at 43.89% and Fish Audio S2 Pro at 43.80%, although Qwen3-TTS-1.7B is higher on complex pronunciation (30.00% against 25.10%). On one H100 in bf16, Boson measured a real-time factor of 0.147 for a single request, with a mean latency of 617 ms on Seed-TTS English sentences, and about 62 seconds of audio per second at 16 concurrent requests. All of these are developer-reported, and the Emergent TTS win rates come from an automated judge.

The weights are one 9.31 GB BF16 file, and the card shows serving through SGLang-Omni or vLLM-Omni with an OpenAI-style speech endpoint and streaming. The license is research and non-commercial, and Boson calls it source-available rather than open source. A Creator Use Grant lets digital creators publish and monetize podcasts, videos, audiobooks and social posts with the model's audio for free if they credit Boson AI's Higgs Audio, but production use, hosting, embedding in a product and training other speech models all need a separate agreement. In the 30 days to 11 November 2026, Hugging Face counted about 114,000 downloads.

Good for

  • Expressive conversational speech where you place laughter, whispers, pauses and emotions in the text
  • Reading in lower-resource languages, since Boson reports single-digit error rates in 102 of them
  • Podcasts, videos and audiobooks that you publish on your own channels, which the creator grant lets you monetize for free with credit
  • Research and evaluation of controllable, multilingual speech synthesis

Watch out for

  • It is not open source, and production use, hosted or API use, and embedding it in a product need a commercial license from Boson, which the creator grant does not cover
  • The license bans using the weights or outputs to train other speech or generative models, requires explicit consent to clone a real person's voice and requires AI disclosure where the law demands it
  • The benchmarks and the 102-language claim are Boson's own, the Emergent TTS win rates come from an automated judge, and the model trails Qwen3-TTS-1.7B on complex pronunciation
  • The card gives no minimum hardware, and the measurements are on an H100 with a 9.31 GB BF16 file
  • Sound-effect tags work best next to a written sound such as "Haha", emotion and style tags go at the start of the text, and the context is 8,192 tokens

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Higgs TTS 3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Higgs TTS 3 4B BF16 4.65B 9.31 GB 11 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

Boson AI's license for the Higgs TTS 3 weights, last updated on 8 July 2026. It calls itself a source-available license, not an open-source one. Research and non-commercial use is free, and any commercial purpose needs a separate written license from Boson. The one exception is a Creator Use Grant, which lets digital creators make and monetize podcasts, videos, audiobooks and social posts with the model's audio if they credit Boson AI's Higgs Audio.

  • Attribution or model name must be shown
  • Acceptable-use policy applies

More about Boson Higgs Non-Commercial

See also

Frequently asked questions

What is Higgs TTS 3?
Higgs TTS 3 turns text into expressive speech in more than 100 languages and can copy a voice from a short recording. You can put tags in the text for emotions, whispering, laughter, speed and pauses. The weights are free for research and personal use, and for creators who credit Boson, but a product or hosted service needs a paid license.
Can I use Higgs TTS 3 commercially?
Not without separate permission. Higgs TTS 3 is released under the Boson Higgs Non-Commercial, which allows research and personal use only.
Is Higgs TTS 3 open source?
Higgs TTS 3 is open-weight, not open source in the strict sense. You can download the model files, but the Boson Higgs Non-Commercial adds conditions that the Open Source Initiative definition does not allow.
Can I run Higgs TTS 3 on my own computer?
Yes. The smallest variant, Higgs TTS 3 4B, needs about 11 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Higgs TTS 3 released?
Higgs TTS 3 was released by Boson AI on June 4, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Drama 3 Fish Audio 35.9B 3B active – Fish Audio Community 1.0 (Conditional) 25 GB Gaming GPU
OmniVoice Xiaomi 613M – CC BY-NC (Non-commercial) 3.75 GB Laptop or phone
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
Voxtral TTS Mistral AI 4.09B – CC BY-NC (Non-commercial) 9.2 GB Laptop or phone