IBM has released Granite Speech 5.0 2B, an open-source model designed for transcription and translation across six languages. The system ships under an Apache 2.0 licence and updates two existing tools, the edge-focused Turbo CTC and the diarization-focused 4.1 2B Plus. Developers can now access accurate transcription, rapid on-device processing, and transcripts that identify speakers.
Technical specifications
The model contains 2 billion parameters in a dense architecture. It uses a YaRN extension to handle 16k positions. IBM recommends processing audio segments of roughly two minutes.
Deployment options include GPU via Transformers or vLLM, Apple Silicon via mlx-audio, and local CPU or Mac via llama.cpp GGUF, with Q8_0 quantisation suggested.
On the Open ASR Leaderboard, it ranks second among open-weight models with an average English Word Error Rate (WER) of 5.68%.
Performance peaks at 0.99% WER on the LibriSpeech Clean dataset. Accuracy drops significantly on the leaderboard’s private conversational set, where the WER reaches 13.79%.
Architecture and training
According to the model card, the system consists of three components. A speech encoder made of 16 Conformer blocks trains with Connectionist Temporal Classification (CTC). This block uses attention over four-second audio segments.
A two-layer window q-former then projects and downsamples acoustic embeddings. The total downsampling factor is 10x, meaning the LLM receives embeddings at 10 Hz.
The LLM is an intermediate checkpoint of granite-4.0-1b-base. IBM extended it to 16k positions using YaRN and fine-tuned it with LoRA adapters.
Training consumed 390,000 hours of public audio alongside synthetic data. The synthetic sets targeted Japanese ASR, keyword biasing, and translation. The process took 40 days on 8 H100 GPUs.
Improvements over Granite Speech 4.1 2B
- Language ID prompting: Users must name the input language in the prompt to steer transcription.
- Keyword lists up to 2,000 terms: This allows biasing output toward specific product names, people, and acronyms.
- Punctuation and capitalization on request: Capitalisation F1 scores rise from 90.16 to 99.51 across test sets.
- Robustness: The model handles noise, reverberation, and distant microphones better than its predecessor.
Benchmark performance
IBM’s chart on the Open ASR Leaderboard shows a 5.68% average English WER. This figure covers nine public and two private test sets as of October 21, 2026. Only Qwen3-ASR-1.7B scores better, at 5.40%.
A source conflict exists regarding this figure. IBM’s blog reports 5.06% instead. That number is unofficial, dated October 15, and averages only public short-form sets.
Results vary by dataset. The model scores 0.99% on LibriSpeech Clean and 2.21% on LibriSpeech Other. The private conversational set yields a 13.79% WER.
On the noisy FFASR leaderboard, it ranks fourth out of 42 models at 14.80% average WER. Qwen3-ASR-1.7B scores 13.41% there.
Multilingual CommonVoice WER is 2.60% for Portuguese and 3.50% for Spanish. Japanese is the weakest language at 10.74%.
For translation, X-to-English CoVoST BLEU reaches 53.6 for Portuguese and 40.1 for German. English-to-Mandarin was not evaluated.
Keyword biasing shows the clearest gain. On Earnings-22, keyword F1 rises from 67.2 to 90.1 with correct keyword lists. With 2,000-term lists mixing correct and wrong terms, it still reaches 82.8.
Comparison with other open ASR models
| Feature | Granite Speech 5.0 2B | Qwen3-ASR-1.7B | Cohere Transcribe 03-2026 | Canary-Qwen-2.5B |
|---|---|---|---|---|
| Size | 2B | 2.35B (Hub count) | 2.07B (Hub count) | 2.5B |
| License | Apache 2.0 | Apache 2.0 | Apache 2.0 (gated) | CC-BY-4.0 |
| ASR languages | 6 | 30 + 22 Chinese dialects | 14 | English only |
| Speech translation | Yes | Not disclosed | Not disclosed | No |
| Open ASR avg WER (Oct 21, 2026)* | 5.68% | 5.40% | 5.77% | 5.94% |
| FFASR avg WER (Oct 18, 2026)* | 14.80% | 13.41% | 15.03% | Not disclosed |
| Keyword biasing | Up to 2,000 terms | Not disclosed | Not disclosed | Not disclosed |
| Long audio | ~2 min segments | Long audio supported | Not disclosed | 40 s max in training |
*Leaderboard averages as charted in IBM’s model card from the Open ASR and FFASR leaderboards.
Updates to Turbo CTC and 4.1 2B Plus
Granite Speech 5.0 Turbo CTC maintains its 470M footprint. It now transcribes seven languages with native punctuation and capitalisation. IBM compiled one Conformer block and reused its graph across the stack. That lifted throughput to 19,870 RTFx, up from about 12,600 at launch.
The updated Granite Speech 4.1 2B Plus is a drop-in weight update. Average timestamp shift falls from 210.86 to 98.54. Long-form WER drops from 13.53 to 10.22. Biased-keyword WER falls from 11.79 to 6.48.
Deployment
Granite Speech 5.0 2B runs in Transformers, vLLM, llama.cpp and mlx-audio. For GGUF, IBM recommends Q8_0. Lower-bit quantisations lose accuracy on audio longer than about one minute.
Practical impact
Developers can now run high-quality transcription on local hardware without relying on cloud APIs. The ability to inject up to 2,000 specific keywords means the model can be tuned for earnings calls, legal proceedings, or technical support logs without retraining. The update to Turbo CTC significantly improves speed for edge devices, while the revised 4.1 2B Plus fixes timing issues that previously plagued long-form recordings.
