Models / IBM/ Granite Speech 3.3 2b

Granite Speech 3.3 2b

IBM · released Apr 28, 2025 · ibm-granite/granite-speech-3.3-2b

Speech to textTranscribes a recording into words

Input: audio. Output: text.InputOutput
Type
Open weightsApache License 2.0
Languages
5
Size
3B

Context measured in tokens

Our take

Written Aug 2, 2026

Granite Speech 3.3 2b is an Apache-licensed speech-to-text model from IBM that turns audio into written text across five languages. It is extremely fast and accurate on clean recordings, though its error rate rises sharply on accented speech and podcasts.

Who should pick it

Pick this for batch transcription where speed matters — an hour of audio in roughly seven seconds. Use it for clean read-aloud material or self-hosted deployments needing a permissive licence across five languages. Skip it if your audio is accented, from meetings, or from podcasts and video, where accuracy drops sharply; or if you need a hosted API rather than self-hosting.

The case for it

  • Extremely fast transcription at 507.1 times real time — an hour of audio in approximately seven seconds on benchmark hardware.
  • Apache License 2.0 allows commercial use, modification and redistribution.
  • Strong on clean read-aloud speech: 1.21% word error rate, well below its own overall average.
  • Financial calls at 3.5% word error rate, noticeably better than its 9.03% on general podcasts and video.

The case against it

  • Accuracy collapses on challenging real-world audio: 9.73% word error rate on accented speech, roughly eight times worse than on clean read-aloud.
  • No accuracy data for its four non-English languages; every figure here is English-only.
  • No hosted offers in our catalogue, so self-hosting is the only option.
00

How good is it?

TranscriptionTurning speech into text3 of 5Open ASR WER · 30th of 74

Words it gets right

94.5%

Misses roughly one word in 18, averaged over nine English test sets.

How fast it listens

507×39th of 62

an hour of audio in 7 seconds, on the board's own hardware. Your machine will differ.

Languages

5

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

Where it struggles
Read aloudaudiobooks, clean recording1.2%
Podcasts and videoeveryday internet audio9%
Accented speechspeakers from many countries9.7%
Meetingsa room, several people, far microphone7.8%

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

Which languages

English · French · German · Spanish · Portuguese

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
Recorded meetings 12th of 74Accented speech 21st of 74Clean read speech 22nd of 74Harder read speech 27th of 74Podcasts and video 57th of 74Financial calls 61st of 74European-accented speech 61st 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.
507.1independentsource ↗
5.5independentsource ↗
9.7independentsource ↗
3.5independentsource ↗
7.8independentsource ↗
1.2independentsource ↗
2.8independentsource ↗
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_M1.9 / 24 GBest
Spare memory19.5 GB spare
Usable context131Kwhat the spare memory holds; no published limit on record
Decode speed443 tok/sest

Room to spare. 19.5 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_M1.9 / 32 GBest
Spare memory27.5 GB spare
Usable context262Kwhat the spare memory holds; no published limit on record
Decode speed788 tok/sest

Room to spare. 27.5 GB spare means a 10% error in the size would not change the answer.

On a MacFits in memory

Apple M2 (8-core GPU, 8GB unified) · 8 GB

Weights at Q4_K_M1.9 / 8 GBest
Spare memory2.7 GB spare
Usable context33Kwhat the spare memory holds; no published limit on record
Decode speed39 tok/sest

Room to spare. 2.7 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
1.9 GBest
Fits in memory
Q5_K_M
2.2 GBest
Fits in memory
Q8_0
3.3 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 Granite Speech 3.3 2b

03

When we formed this view

Dates behind this page

Aug 2, 2026BenchmarkScored 9.7 on Accented speechleaderboard
Aug 2, 2026BenchmarkScored 3.5 on Financial callsleaderboard
Aug 2, 2026BenchmarkScored 7.8 on Recorded meetingsleaderboard
Aug 2, 2026BenchmarkScored 4.5 on European-accented speechleaderboard
Aug 2, 2026BenchmarkScored 9 on Podcasts and videoleaderboard
Aug 2, 2026BenchmarkScored 1.2 on Clean read speechleaderboard
Aug 2, 2026BenchmarkScored 2.8 on Harder read speechleaderboard
Aug 1, 2026ListedListed on LLMapfirst indexed by our pipeline
Aug 1, 2026BenchmarkScored 507.1 on Open ASR RTFxleaderboard
Aug 1, 2026BenchmarkScored 5.5 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
ibm-granite-granite-speech-3-3-2b

Machine-readable model card (omc.json) →

Something wrong on this page? Tell us