Granite Speech 5.0 enables more accurate and faster transcription, translation, and attribution.
IBM Granite Speech 5.0 adds accurate multilingual transcription, faster Turbo CTC, and improved speaker attribution under Apache 2.0.
Key Takeaways
- IBM released three Granite Speech models: granite-speech-5.0-2b for accurate multilingual transcription and translation, granite-speech-5.0-470m-turboctc for fast edge processing, and granite-speech-4.1-2b-plus for speaker-aware long meetings or calls.
- granite-speech-5.0-2b improves English noise robustness, supports language prompts, up to 2,000 keyword terms, and formatted output, with a reported average English WER of 5.06% on the Open ASR Leaderboard.
- granite-speech-5.0-470m-turboctc adds multilingual transcription and native punctuation/capitalization while increasing throughput to 19,870 RTFx, about 1.6x faster than its launch version.
- granite-speech-4.1-2b-plus is a drop-in upgrade that improves speaker attribution, timestamps, and long-form transcription, cutting long-form WER by 24% and timestamp shift by more than half.
- All three models are Apache 2.0 licensed, supported in Transformers, and IBM recommends pairing them with Granite Guardian for production risk detection.
IBM has released three updates to its Granite Speech collection: granite-speech-5.0-2b, a revised granite-speech-5.0-470m-turboctc, and an updated granite-speech-4.1-2b-plus. The update matters most for teams building automatic speech recognition systems, because the three models split the work into accurate transcription and translation, fast edge processing, and speaker-aware handling of long meetings or support calls. The models are released under an Apache 2.0 license, and the main model’s benchmark figures are dated 21 October 2026.
granite-speech-5.0-2b targets accuracy and language control
This model succeeds granite-speech-4.1-2b at the same 2B parameter count. It covers transcription in English, French, German, Spanish, Portuguese, and Japanese. For translation, it handles those languages to and from English, plus English to Italian and Mandarin. The model was trained on 390,000 hours of public audio, with synthetic data built for Japanese automatic speech recognition, keyword biasing, and translation.
Compared with the earlier 4.1 2B Speech model, the update adds five practical changes. English transcription is reported as more accurate on short-form benchmarks and more resilient to noise, reverberation, and distant-microphone speech. Users can also name the input language in the prompt to steer multilingual transcription. Translation quality is aimed at longer, more natural conversational audio, not only short read sentences. Keyword lists can contain up to 2,000 terms, biasing transcription and translation toward product names, people, and acronyms in each supported language. Finally, formatted output with punctuation and capitalization can be requested in the prompt; capitalization F1 scores, which balance precision and recall, range from 89.7 to 99.5 across the test sets used for the release.
The model is available under Apache 2.0 with support in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon. The release cites official Open ASR Leaderboard results dated 21 October 2026: granite-speech-5.0-2b ranks #2 among top open-weight automatic speech recognition models on average English word error rate (WER), with 5.06%. WER is an error rate, so lower is better.
Turbo CTC adds multilingual work and faster throughput
granite-speech-5.0-470m-turboctc first appeared in August as an English-only, 470M-parameter encoder built for speed. The revised version keeps that size while adding the multilingual transcription, native formatting, and robustness users asked for. It now transcribes English, Japanese, Spanish, German, French, Italian, and Portuguese. Transcripts arrive with punctuation and capitalization already applied, removing the need for a separate punctuation or truecasing step. The release says the training also targets reverberant audio, improving behavior in echoey rooms and far-field recordings.
The speed gain comes from compiling a single Conformer block and reusing its graph across the stack, which reduces warmup time. On the Open ASR Leaderboard, the revised model reaches 19,870 RTFx, up from roughly 12,600 RTFx at launch. That is about 1.6x faster at the same accuracy level. RTFx is a throughput metric, and higher values are better. The chart uses average WER against RTFx on nine public English short-form test sets, with results dated October 2026.
granite-speech-4.1-2b-plus focuses on who spoke and when
The updated granite-speech-4.1-2b-plus is aimed at recordings where speaker identity and timing matter: meetings, support calls, and interviews with multiple speakers over long periods. It keeps the 4.1 name because the architecture, prompt format, and API are unchanged. That makes it a drop-in replacement for the current model, with improved weights.
The revision improves four areas. Speaker attribution is more accurate, with word diarization error rate (WDER) dropping 7% and speaker-attributed WER (cpWER) dropping 7% overall. Timestamp quality improves as average timestamp shift (AAS) is cut by more than half, helping captions, search, and review. Long-form transcription becomes stronger, with WER on long-form test sets dropping 24%. Keyword biasing also improves: errors on biased keywords fall 45% when using 20-term keyword lists. As with the other models, all these metrics are error rates, so lower values are better.
Model choice and next steps
The lineup gives teams a clearer selection path. Use granite-speech-5.0-2b when high transcription accuracy, translation, or large keyword lists matter. Use granite-speech-5.0-470m-turboctc when raw speed or on-device transcription is the priority, and try it in the browser with IBM’s WebGPU streaming demo. Use granite-speech-4.1-2b-plus for meetings or multi-speaker calls where labels and timing are central.
All three models are Apache 2.0 licensed and have native support in Transformers. Full evaluation results, architecture details, training data, and usage examples are published on each model card. IBM recommends pairing these speech models with Granite Guardian for production deployments that need additional risk detection.
| Model | Strength | Key figure |
|---|---|---|
| granite-speech-5.0-2b | Accurate transcription, translation, and keyword control | Average English WER 5.06%; keyword lists up to 2,000 terms |
| granite-speech-5.0-470m-turboctc | Fast, multilingual transcription for edge work | 19,870 RTFx on Open ASR Leaderboard |
| granite-speech-4.1-2b-plus | Speaker attribution, timestamps, and long speech | 24% lower long-form WER |

Comments (0)