whatmodel.sh / tts / Higgs Audio V2

Higgs Audio V2

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Higgs Audio V2 is an open-weights text-to-speech model from Boson AI, released July 2025. Its published size is 5.8B parameters. Licence: boson-higgs-audio-2-community.

Catalogue identity notes

Higgs Audio V2 as explicitly linked in the result row.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

5.8 = the maker's own stated total '3.6B (LLM) + 2.2B (Audio Dual FFN)' (the card writes the sum, not the result). Excludes the separate audio tokenizer (bosonai/higgs-audio-v2-tokenizer). Note the repo name says 3B (the Llama-3.2-3B base).

54.7±3.1#53 of 118 text-to-speech modelsfrom 7 benchmark results across 3 sources

Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.

$ how it compares

The text-to-speech models ranked either side of Higgs Audio V2, out of 148 with a score, then more from Boson AI.

#modelscorewhereRAMreleased
64Supertonic 3 Supertone55±10local~0.4GB2026-05
65ElevenLabs v3 Alpha ElevenLabs55±9comparison only2025-06
66Higgs Audio V2 Boson AI55±3local~7GB2025-07
67Fasih-TTS-V1 NightPrince (Yahya Elnawasany)55±10local~1.9GB2026-07
68CosyVoice-300M Alibaba (FunAudioLLM)54±3local~2.3GB2024-07
more from Boson AI
23Higgs Audio V3 TTS Boson AI65±3api $15.00/M chars +local~10GB2026-06

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

voice cloning▁·▁▁▁▁▁▁▇█▃▃ #26 of 37 * · 32% of its score · 1 test

Beats CosyVoice2-0.5B, MaskGCT here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Seed-TTS-Eval · test-en/test-zh macro mean error lab2.1% ✎▄··▄▄····█▄▄ beats 0/6avg 3.7%▇▇▇▇▇▇▇▇▇▇▇▇▇ 32%typical
other tests6 not in a task · 64% of its score
testresultvs the others in its rundifficultycountswhy
EmergentTTS · 6-category preference lab38.9% ✎▅▅▃▃█▅▅▃·▃·▅ beats 12/16hard · avg 30%▇▇▇▇▇▇▇▇▇▇▇▇ 31%at its level
RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity56.0% ✎▃·····▂▂▂▃█▆ beats 9/15medium · avg 46%▇▇▇▇▇▇ 14%blunt · at its level
EmergentTTS · 6-category WER lab19.4% ✎▂▄··▂▄▄··█▄▆ beats 4/16avg 16%▇▇▇ 7.2%blunt
MiniMax multilingual test set · 23-language mean error lab49.9% ✎▃·▃·▃·····▃█ beats 1/6avg 44%▇▇▇ 6.9%sharp · noisy
CV3-Eval multilingual voice clone · 9-language mean error lab21.2% ✎▃·····▃···▅█ beats 1/6avg 20%▇▇ 5.0%noisy
RVCBench LibriTTS clean voice cloning · Whisper-medium WER25.0% ✎▂▃▂··▂·▃▃▃█▂ beats 1/15avg 8.6%▇ 0.2%blunt · noisy

$ where to run it

Other ways to run it

version

  1. Run it as published · needs ~12.37 GB, NVIDIA only · from huggingface.co/bosonai (official)
    uv run --with 'transformers>=5.3.0' --with torch --with accelerate python -c "from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration; model_id = 'bosonai/higgs-audio-v2-generation-3B-base'; processor = AutoProcessor.from_pretrained(model_id, device_map='auto'); model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map='auto'); conversation = [{'role': 'system', 'content': [{'type': 'text', 'text': 'Generate audio following instruction.'}]}, {'role': 'scene', 'content': [{'type': 'text', 'text': 'Audio is recorded from a quiet room.'}]}, {'role': 'user', 'content': [{'type': 'text', 'text': 'The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.'}]}]; inputs = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=True, return_dict=True, sampling_rate=24000, return_tensors='pt').to(model.device); outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=False); decoded = processor.batch_decode(outputs); processor.save_audio(decoded, 'output_single_speaker.wav')"

Official weights and documentation: huggingface.co/bosonai ↗

$ how the score was worked out · Higgs Audio v3 comparison tab… 44%, EmergentTTS-Eval six-category… 38%, RVCBench: Benchmarking Robust… 14%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Higgs Audio v3 comparison tables in SGLang-Omni cookbook (Boson AI contribution) · lab-reported3▇▇▇▇▇▇▇▇▇▇▇▇▇ 44%Apache-2.0
EmergentTTS-Eval six-category leaderboard · lab-reported2▇▇▇▇▇▇▇▇▇▇▇ 38%Apache-2.0
RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2)2▇▇▇▇ 14%CC BY 4.0
prior · pull toward the middle▇ 4%
weighted average= 54.7

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json