IBM Granite Speech 5.0 2B ranks #2 among open ASR models at 5.06% WER under Apache 2.0

IBM released granite-speech-5.0-2b, a 2-billion-parameter speech-language model that, according to IBM, ranks #2 among open-weight ASR models with an average English word error rate (WER) of 5.06%. The company also shipped revised versions of two existing models: granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. All three are released under Apache 2.0.

IBM's ranking is unofficial, dated 15 October 2026, and averaged over the Open ASR Leaderboard's public English short-form test sets.

granite-speech-5.0-2b

The model succeeds granite-speech-4.1-2b at the same parameter count. IBM says it was trained on 390,000 hours of public audio, plus synthetic data built for Japanese ASR, keyword biasing, and translation.

• Transcription: English, French, German, Spanish, Portuguese, Japanese.
• Translation: to and from English in those languages, plus English to Italian and Mandarin.
• Language-ID prompting: naming the input language in the prompt steers multilingual transcription.
• Keyword biasing: lists of up to 2,000 terms, in every supported language.
• Formatting on request: punctuation and capitalization via prompt; IBM reports capitalization F1 of 89.7 to 99.5 across its test sets.
• Robustness: IBM claims better handling of noise, reverberation, and distant-microphone speech, and better long-form and conversational translation.

Support is available in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon. Context window, pricing, and training cutoff date were not disclosed. As an open-weight release, there is no listed API price.

granite-speech-5.0-470m-turboctc

Introduced in August as an English-only, 470M-parameter encoder, Turbo CTC now supports English, Japanese, Spanish, German, French, Italian, and Portuguese. It outputs native punctuation and capitalization, with no separate truecasing step. IBM says training that explicitly targets reverberant audio improves performance in echoey and far-field recordings.

IBM compiled a single Conformer block and reused its graph across the stack, which it says cut warmup time. The model now reaches 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch. IBM puts that at about 1.6x faster at the same accuracy level. A WebGPU streaming demo is available for in-browser testing.

granite-speech-4.1-2b-plus

This revision keeps the 4.1 name because architecture, prompt format, and API are unchanged, making it a drop-in weight replacement. It targets multi-speaker recordings. IBM's reported results (all error rates, lower is better):

Metric	Previous	Updated
Timestamps, all speech (AAS)	210.86	98.54
Timestamps, short speech (AAS)	29.27	26.20
Keyword biasing, 20-term lists (biased-word WER)	11.79	6.48
Long-form WER	13.53	10.22
Speaker attribution (WDER)	4.40	4.10
Speaker-attributed WER (cpWER)	30.50	28.48

These figures come from IBM's own evaluations. Full results are on the model cards.

What this means

IBM is splitting its speech line by deployment constraint rather than chasing one leaderboard number. The 2B model targets accuracy and control, Turbo CTC targets throughput and on-device use, and the 4.1 Plus weights target diarization-heavy enterprise audio.

The largest gains are in the 4.1 Plus update. Halving timestamp error and cutting long-form WER by 24% matter more for meeting and call-analysis pipelines than a small gain in short-form WER. The speaker-attribution improvement is modest, at about 7% on both WDER and cpWER.

The #2 open-weight ranking rests on English short-form benchmarks and IBM's own snapshot. Teams working in noisy conversational audio or non-English languages should run their own evaluations. The Apache 2.0 license and broad runtime support (vLLM, llama.cpp, mlx-audio) lower the barrier to doing so.