Models / Boson AI/ Higgs Audio v3 8b STT v2

Higgs Audio v3 8b STT v2

Boson AI · released Apr 27, 2026 · bosonai/higgs-audio-v3-8b-stt-v2

Speech to textTranscribes a recording into words

Input: audio. Output: text.InputOutput
Type
Open weightsApache License 2.0
Languages
1
Size
8.9B

Context measured in tokens

Our take

Written Aug 2, 2026

Higgs Audio is an English-only speech-to-text model from Boson AI that turns audio into written words at exceptional speed. It achieves near-perfect accuracy on clean read-aloud recordings, but its error rate rises sharply on accented speech or meeting audio.

Who should pick it

Pick this for high-throughput transcription of clean English audio where speed matters, or for self-hosted deployment needing a permissive licence. Skip it if your audio includes accented speakers or meetings, if you need non-English languages, or if you require hosted inference.

The case for it

  • Extremely fast: 139.1 times real time on benchmark hardware, or roughly an hour of audio in 26 seconds.
  • Near-perfect on clean read-aloud English: 0.95% word error rate, roughly one word wrong per 105.
  • Apache 2.0 licence allows commercial use, modification and redistribution.

The case against it

  • Accuracy collapses in challenging conditions: 8.45% on accented speech and 8.37% in meetings — roughly eight to nine times worse than on clean read-aloud.
  • No hosted availability in our catalogue; must be self-hosted or sourced outside our data.
  • English only; no source we hold measures accuracy in any other language.
00

How good is it?

TranscriptionTurning speech into text4.5 of 5Open ASR WER · 11th of 74

Words it gets right

95.3%

Misses roughly one word in 21, averaged over nine English test sets.

How fast it listens

139×55th of 62

an hour of audio in 26 seconds, on the board's own hardware. Your machine will differ.

Languages

1

Listed on the model card. The accuracy above is English only.

Where it struggles
Read aloudaudiobooks, clean recording1%
Podcasts and videoeveryday internet audio7.1%
Accented speechspeakers from many countries8.5%
Meetingsa room, several people, far microphone8.4%

Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case, not to the board.

The figures above come from the Open ASR Leaderboard, an independent public test that runs every model on the same recordings. It is the only measurement of transcription quality we know of, so there are no other scores to show.

Also scored, on boards we give no mark for
Harder read speech 3rd of 74Podcasts and video 4th of 74Clean read speech 4th of 74Accented speech 11th of 74European-accented speech 19th of 74Recorded meetings 22nd of 74Financial calls 55th of 74

These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.

Every published score for this model9 scoresEvery figure we hold, from 9 boards, with who ran it and a link to the source — including the boards no rating above is built on.
139.1independentsource ↗
4.7independentsource ↗
8.5independentsource ↗
3.2independentsource ↗
8.4independentsource ↗
7.1independentsource ↗
0.95independentsource ↗
2.1independentsource ↗
01

Can you run it yourself?

Fits in memory
weights load entirely on the card
Spills to system RAM
some weights offload; much slower
Too large
will not load even with offload
est
size is calculated; the verdict could change by 10%

Comfortable fit

A card many people ownFits in memory

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M5.6 / 24 GBest
Spare memory15.5 GB spare
Usable context66Kwhat the spare memory holds; no published limit on record
Decode speed150 tok/sest

Room to spare. 15.5 GB spare means a 10% error in the size would not change the answer.

One step upFits in memory

GeForce RTX 5090 · 32 GB

Weights at Q4_K_M5.6 / 32 GBest
Spare memory23.5 GB spare
Usable context131Kwhat the spare memory holds; no published limit on record
Decode speed266 tok/sest

Room to spare. 23.5 GB spare means a 10% error in the size would not change the answer.

On a MacFits in memory

Apple M1 (8-core GPU) · 16 GB

Weights at Q4_K_M5.6 / 16 GBest
Spare memory4.7 GB spare
Usable context33Kwhat the spare memory holds; no published limit on record
Decode speed9 tok/sest

Room to spare. 4.7 GB spare means a 10% error in the size would not change the answer.

Your hardware
Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M
recommended
5.6 GBest
Fits in memory
Q5_K_M
6.6 GBest
Fits in memory
Q8_0
9.8 GBest
Fits in memory

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

02

Models people weigh against Higgs Audio v3 8b STT v2

03

When we formed this view

Dates behind this page

Aug 2, 2026BenchmarkScored 8.5 on Accented speechleaderboard
Aug 2, 2026BenchmarkScored 3.2 on Financial callsleaderboard
Aug 2, 2026BenchmarkScored 8.4 on Recorded meetingsleaderboard
Aug 2, 2026BenchmarkScored 2.9 on European-accented speechleaderboard
Aug 2, 2026BenchmarkScored 7.1 on Podcasts and videoleaderboard
Aug 2, 2026BenchmarkScored 0.95 on Clean read speechleaderboard
Aug 2, 2026BenchmarkScored 2.1 on Harder read speechleaderboard
Aug 1, 2026ListedListed on LLMapfirst indexed by our pipeline
Aug 1, 2026BenchmarkScored 139.1 on Open ASR RTFxleaderboard
Aug 1, 2026BenchmarkScored 4.7 on Open ASR WERleaderboard

Prices last checked 6h ago

What we do not know about this model yet

  • We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
  • Nothing we hold says whether an endpoint streams, so we do not show it either way.
04

Licence and identifiers

What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

Apache License 2.0

permissiveCommercial use allowed

Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.

Identifiers

Architecture
Dense
Modality record
audio->text
Catalogue slug
bosonai-higgs-audio-v3-8b-stt-v2

Machine-readable model card (omc.json) →

Something wrong on this page? Tell us