Two averages, both published by IBM

The 5.06% figure is the one that will travel. It is in IBM's blog post, dated 21 October, under a chart, with a clean ranking attached: second among top open-weight models on average English word error, results official as of 21 October 2026. The chart's own caption adds the scope. That average covers the Open ASR Leaderboard's public English short-form test sets. Short-form, public, and read or prompted speech feature heavily in that kind of set.

The model card, dated a day earlier, tells a wider story. Its leaderboard chart, titled for granite-speech-5.0-2b itself, averages nine public and two private English test sets, and comes out at 5.68%. Neither number is a typo, and neither is independent: both are IBM's, computed over different denominators, and the per-set values printed in the card's chart are arithmetically consistent with both averages. The ranking survives the change of scope. Qwen3-ASR-1.7B still leads on the fuller average, at 5.40%, so Granite Speech is second on either count. What changes is how close second place looks, and how much of the picture the two private sets add.

The spread inside the average is the usable news

Averages over unlike audio conceal the thing a buyer needs to know, which is how a model handles their audio. Granite Speech 5.0 scores 0.99% word error on LibriSpeech Clean, the cleanest read-speech set in the group. It scores 13.79% on the leaderboard's private conversational set. That is not a contradiction; it is a description of two different jobs. Reading an audiobook aloud in a quiet room and transcribing overlapping conversation are both called speech recognition. The two measured error rates, 0.99% against 13.79%, differ by a factor of almost fourteen in measured error. That does not mean conversational transcription is fourteen times more difficult. It means a result measured on clean read speech may not predict what the same model does with real conversation.

The noisy-audio results point the same way. On the FFASR leaderboard, which tests reverberant and far-field speech, Granite Speech 5.0 ranks fourth of 42 at 14.80% average word error, again behind Qwen3-ASR-1.7B at 13.41%. The language spread is also wide: on CommonVoice, Portuguese scores 2.60% and Spanish 3.50%, while Japanese, the language IBM built synthetic data specifically to support, is the weakest of the six at 10.74%. A team evaluating Portuguese dictation and a team evaluating Japanese customer calls may see very different error profiles from the same model. The CommonVoice results illustrate how much language variation there is inside one model; they do not directly predict performance on either team's own recordings.

What is genuinely useful in this release

Strip the ranking away and the practical offer is solid. The model is small, openly licensed under Apache 2.0, and runs in the stacks people already use for local inference, including llama.cpp on a laptop and mlx-audio on Apple Silicon. Running it locally can keep recordings off a hosted API, provided the complete runtime and its integrations are configured to remain local; that is a property of the deployment, not of the weights alone. Two controls are built into the prompting: users can name the expected input language, and they can supply keyword lists. Keyword biasing is the feature with the clearest measured gain. On Earnings-22, the model's keyword F1 — its score for recognising specified keywords, with common words excluded from the evaluation — rose from 67.2 without biasing to 90.1 with a correct list applied at inference time, and even a 2,000-term list mixing correct and incorrect terms reached 82.8. That measures how well the listed terms come through, not overall transcription accuracy. For anyone whose transcripts keep mangling the same product names, it is worth more than a decimal point of average error. IBM also evaluates punctuation and capitalisation separately, under their own metrics; those scores are not part of the word error rates above.

IBM is also unusually plain about the operating limits, and its guidance is best read as IBM's recommended configurations rather than universal limits. Transcribe in segments of about two minutes or less, and split longer recordings before decoding. If you quantise the GGUF build, use Q8_0, which IBM says matches the unquantised model in its tests; lower-bit versions lose accuracy on audio longer than about a minute. The two companion updates follow the same pattern. Turbo CTC is a speed model that now handles seven languages and reaches 19,870 RTFx on IBM's reported Open ASR Leaderboard evaluation, up from roughly 12,600 at launch; RTFx is a throughput multiple that depends on workload, hardware and batching, not a promise about any particular machine. The revised 4.1 2B Plus is a drop-in weight update for meetings, cutting average timestamp shift from 210.86 to 98.54 and long-form error from 13.53% to 10.22%. Test Granite Speech on an hour of your own meetings before you test it on LibriSpeech. IBM has published enough of its own numbers to tell you which result to expect.

Sources

  1. [1] IBM Granite team, “Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster”, Hugging Face blog, 21 October 2026 (primary source for the 5.06% average English WER and #2 open-weight ranking, results official as of 21 October 2026 and averaged over the Open ASR Leaderboard's public English short-form test sets, the six transcription languages and translation pairs, Apache 2.0, 390,000 training hours, language-ID prompting, keyword lists up to 2,000 terms, Turbo CTC at 19,870 RTFx on the Open ASR Leaderboard up from roughly 12,600 at launch, and the 4.1 2B Plus table: average timestamp shift 210.86 to 98.54, long-form WER 13.53% to 10.22%)Read source
  2. [2] IBM Granite, granite-speech-5.0-2b model card, Hugging Face, release date 20 October 2026 (primary source for the card's English WER chart: 5.68% average over nine public and two private English test sets, Qwen3-ASR-1.7B first at 5.40%, per-set scores including 0.99% on LibriSpeech Clean and 13.79% on the private conversational set; the FFASR chart: #4 of 42 at 14.80%, Qwen3-ASR-1.7B at 13.41%; the CommonVoice chart: Portuguese 2.60%, Spanish 3.50%, Japanese 10.74%; the keyword-F1 chart and its definition — F1 of transcribed keywords excluding common words — with Earnings-22 at 67.2 without biasing, 90.1 with a correct list and 82.8 with a 2,000-term mixed list; the separate punctuation and capitalisation evaluation; the supported runtimes; the guidance to transcribe in segments of up to about two minutes; the Q8_0 recommendation and the warning that lower-bit quantisations lose accuracy on audio longer than about a minute; and training completed in 40 days on 8 H100 GPUs)Read source
  3. [3] MarkTechPost, “IBM AI Releases Granite Speech 5.0 2B for Multilingual Transcription and Translation”, 22 October 2026 (the write-up that first flagged the 5.06% / 5.68% discrepancy in print; every figure it extracted from the model-card charts was re-verified for this article against IBM's card and blog directly)Read source