Granite Speech 5.0 2B: IBM's New ASR Model Sat on Hugging Face for Two Weeks Before IBM Announced It

The repository for Granite Speech 5.0 2B was created on 7 October 2026. The weights were committed the same evening, under a commit message that says only "Upload granite-speech-5.0-2b checkpoint." The model card then sat there for thirteen days being edited in public — benchmark charts added on the 13th, a release date and official leaderboard results stamped on the 19th, a model signature file on the 20th — before IBM published the blog post that actually announced it, on 21 October 2026. That blog post is one day old as I write this, and it is still the only thing IBM has said about the model anywhere: there is no press release, no ibm.com announcement page, and nothing on the Granite product site beyond the generic landing page. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

So the honest description is not "unannounced" and not "launched with fanfare" either. It is a model that was quietly downloadable for two weeks, whose card was written in the open, and which got its announcement a day before this article. What follows separates the parts that are checkable from the parts that are still just IBM's word.

What is actually in the checkpoint

Granite Speech 5.0 2B is a speech-language model: a Conformer acoustic encoder feeding a small language model through a projector. It is the direct successor to granite-speech-4.1-2b, which shipped on 29 April 2026 at the same parameter count and the same Apache 2.0 licence.

• Size and format — 2B parameters, BF16 safetensors, Apache 2.0, no gating on the repository.

• Encoder — 16 Conformer blocks trained with CTC and two classification heads: 348 character outputs (the first 256 ASCII entries, plus a 92-character phonetic Katakana set for Japanese) and 100,353 BPE outputs drawn from the Granite 4.0 tokenizer. Hidden size 1,024, 8 attention heads of 128, convolution kernel 15, four-second block attention.

• Projector — a two-layer window query transformer over blocks of fifteen 1,024-dimensional acoustic embeddings, with three trainable queries per block and layer. It downsamples 5× on top of the encoder's 2×, so the language model sees 10 Hz embeddings instead of 100 Hz frames.

• Language model — an intermediate checkpoint of granite-4.0-1b-base, extended to 16k positions with YaRN, with LoRA adapters trained jointly with the projector.

• Training — roughly 390,000 hours of public and synthetic audio, 40 days total on 8 NVIDIA H100s (26 days of encoder training, 14 of projector training) on IBM's Blue Vela cluster.

• Languages — English, French, German, Spanish, Portuguese and Japanese for transcription; translation to and from English for those six, plus English into Italian and English into Mandarin.

The 390,000 hours is the number that matters most for what the model is good at. The card breaks it down and the largest single block is not clean read speech: LibriHeavy supplies 45,600 hours of unpunctuated audio and another 45,800 of punctuated audio, VoxPopuli Granary adds 24,000 plus 23,900 punctuated, and then there are 59,900 hours of what IBM calls contextual ASR — the same corpora re-cut with a noisy previous-segment context prepended to the prompt. Another 7,100 and 16,500 hours are keyword-biased ASR in the European and Japanese language groups respectively. That mix is why the headline improvements are punctuation, capitalisation, robustness and keyword biasing rather than raw word error rate on clean English.

The 5.06% number, and the fine print under it

IBM's headline claim is a 5.06% average English word error rate across the nine public short-form test sets of the Open ASR Leaderboard, which it describes as ranking #2 among top open-weight ASR models. The model card publishes the same figure with a different label — "official results as of October 20, 2026" — while the announcement blog calls the underlying run "unofficial, as of 15 October 2026." Those two descriptions are five days and one adjective apart, and nobody outside IBM has reproduced either.

The comparison chart the card carries is worth reading closely, because its selection rules do a lot of work. It shows the top ten open-weight models on average WER, one per organisation, with Granite Speech 5.0 2B placed second:

• #1 Qwen3-ASR-1.7B-hf — 4.94% average WER.

• #2 Granite Speech 5.0 2B — 5.06%.

• #3 cohere-transcribe-03-2026 — 5.21%.

• #4 Hojo-ASR-V1 and #5 higgs-audio-v3-stt — 5.22% each.

• #6 thewhisper-large-v3-turbo — 5.24%. Note that this is a community checkpoint that borrows Whisper's name, not O​penAI's whisper-large-v3-turbo, which scores 7.83% on the same leaderboard. The two are routinely confused and the confusion flatters whichever comparison you are reading.

• #7 canary-qwen-2.5b — 5.38%; #8 ARK-ASR-0.6B — 5.44%; #9 MOSS-Transcribe-Diarize — 5.54%; #10 Voxtral-Small-24B-2507 — 5.81%.

Two things follow. First, the gap to the leader is 0.12 percentage points, which on nine test sets is inside the noise you get from a different decoding configuration — this is a tie for first, described as second place. Second, the spread from #2 to #6 is 0.18 points. Whatever separates these models on clean English, it is not accuracy. IBM's real pitch has to be somewhere else, and it is.

Where the gain is large enough to matter: noise and distance

The FFASR leaderboard measures noisy and reverberant speech — near-field, high SNR, mid SNR and low SNR conditions averaged together. This is the evaluation where the 5.0 generation separates from the 4.1 generation by a margin you can feel, and it is the most useful number in the release.

• Granite Speech 5.0 2B — 14.80% average WER, ranked #4 of 42 models on the board (results as of 18 October 2026).

• granite-speech-4.1-2b-nar — 15.16%, #6.

• granite-speech-4.1-2b — 16.35%, #9.

• For scale: whisper-large-v3 sits at 18.02% on the same board, and nvidia/parakeet-tdt-0.6b-v2 at 17.66%.

That is a 1.55-point improvement over the immediate predecessor on far-field and noisy audio — roughly a 9.5% relative reduction in errors — while the clean-English average moved from 5.33% to 5.06%, about 5% relative. In other words, the improvement is roughly twice as large in the conditions where people actually deploy transcription: conference rooms, call centres, laptop microphones, distant tablet mics. If your ASR stack is already fine on studio audio and falls apart in a meeting room, this release is aimed at you.

Punctuation improved; capitalisation did not

The one numeric table in the card compares 5.0 against 4.1 on punctuation error rate (PER, lower better) and capitalisation F1 (higher better), computed on Levenshtein-aligned word pairs rather than whole sentences. Read the direction of each row, because the picture is not uniform.

• LScln — PER 25.70 → 23.20; Cap-F1 89.71 → 90.16.

• LSoth — PER 22.27 → 21.22; Cap-F1 91.26 → 91.65.

• VoxPopuli — PER 24.86 → 23.70, but Cap-F1 95.35 → 95.07.

• Earnings-22 — PER 22.87 → 21.64, but Cap-F1 95.19 → 94.98.

• CV-EN — PER 9.13 → 9.18 (slightly worse); Cap-F1 96.75 → 96.96.

• CV-DE — PER 3.66 → 3.65; Cap-F1 99.50 → 99.51 (German capitalises nouns, which is why this row is near-perfect).

• CV-ES — PER 11.61 → 11.46, Cap-F1 95.68 → 95.62.

• CV-FR — PER 11.00 → 11.43 (worse); Cap-F1 97.25 → 97.20.

• CV-PT — PER 7.86 → 7.53; Cap-F1 98.51 → 98.48.

Punctuation moved in the right direction on seven of nine sets; capitalisation moved in the wrong direction on five of nine, by between 0.03 and 0.28 points. IBM's summary is "higher transcription accuracy" and "punctuation and capitalisation," which is defensible on the aggregate and slightly generous on the detail. The practical reading: expect fewer missing commas, do not expect the model to suddenly get proper nouns and sentence starts right.

Two features that change integration work

The capabilities most likely to affect an actual deployment are not benchmarked in the card at all, because they are control features rather than accuracy features.

Language-ID prompting. You can prepend "The speech is in German. Transcribe it exactly; do not translate." and the model will transcribe German rather than translating it into English. This matters because a speech model that auto-translates is unusable for, say, a German-language support archive — you get an English transcript and lose the original wording. The prompt also works when written in the target language, which is the more realistic case for a multilingual pipeline.

Keyword biasing up to 2,000 terms. The predecessor's limit is not stated as a hard number in the 4.1 card, but 5.0 raises the supported list to two thousand terms, per prompt, across ASR and speech translation. For product names, drug names, ticket IDs or a customer's own SKU list, this is the difference between a usable transcript and one that needs correction. The card reports keyword-biasing F1 as a chart with and without biasing applied, but no numeric values in text, so treat the size of the improvement as unquantified.

There is also a safety design detail worth knowing: if the model receives a malformed or unfamiliar prompt, IBM says it ignores the prompt and falls back to plain transcription. That is a deliberate constraint on how much audio can influence the system, and IBM recommends pairing the model with Granite Guardian for risk detection in production.

Running it

Granite Speech 5.0 2B is supported natively in transformers>=4.52.1 via AutoModelForSpeechSeq2Seq and AutoProcessor. The card also documents vLLM, llama.cpp and mlx-audio, which matters more than it sounds: this is a model you can run on a laptop, not only behind a GPU cluster.

• Segment length — transcribe in chunks of up to about two minutes and split longer recordings with VAD. The 16k YaRN context is not a licence to feed it an hour of audio; accuracy degrades well before the context does.

• llama.cpp — the card recommends Q8_0 and states that lower-bit quantisation loses accuracy on audio longer than roughly a minute. That is an unusually specific warning and worth taking literally.

• mlx-audio — 0.5.9 or later, plus jinja2, which mlx-audio needs for the chat template but does not install itself.

• Prompts are task-specific and matter — "transcribe the speech with proper punctuation and capitalization" is a different mode from the raw-transcript prompt, and the keyword-biased mode takes a "Keywords: …" suffix.

What is not knowable yet

Three things are worth stating plainly, because the surrounding coverage will not.

First, there is no independent evaluation. Every number above comes from IBM's own eval run or from leaderboard submissions IBM made. The FFASR chart says "official results," which means the submission is on the board, not that a third party ran the model. Nobody has published a reproduction, a failure-mode writeup, or a head-to-head on audio IBM did not choose.

Second, there is no hosted endpoint. The model card states directly that the model "isn't deployed by any Inference Provider," and I could not find it in the OrcaRouter catalogue either — the 209-model listing contains no Granite Speech checkpoint of any generation. Today your options are the weights, or nothing. That is normal for a model that is two days old and it will change; plan for self-hosting in the meantime.

Third, the licences differ across the family in a way that is easy to trip over. The 2B is Apache 2.0 with no commercial restriction. The small TurboCTC model ships in two variants, one Apache 2.0 and one non-commercial — the naming does not make which is which obvious at a glance.

Where this leaves a speech pipeline

The realistic architecture for a Granite Speech 5.0 2B deployment in late 2026 is a self-hosted encoder-decoder transcribing audio, and a separate language-model call doing whatever comes next: summarising the call, extracting fields, classifying intent, cleaning up the transcript against a schema. That second stage is ordinary text generation, and it is the part that benefits from not being hard-wired to one vendor — the transcript goes to one O​penAI-compatible endpoint that can reach 200+ models at provider list price with no markup added, so swapping the cleanup model is a config change rather than a rebuild, and a provider outage fails over instead of taking the pipeline down with it.

For the speech stage itself, the decision this week is narrower than the benchmark chart suggests. Granite Speech 5.0 2B is a credible top-three open-weights English ASR model, a clear improvement on its own predecessor in exactly the conditions that break ASR in production, and a genuinely multilingual model with translation and keyword control that most competitors at this size do not offer. It is also two days old, vendor-benchmarked, and unreproduced. Run it on your own hardest audio before you believe 5.06%.