IBM has released Granite Speech 5.0 2B, a new open speech-language model for transcription and translation. Granite Speech 5.0 2B is a 2B-parameter model from IBM that turns speech into text in 6 languages. It ships under Apache 2.0. IBM also updated 2 existing models: the edge-focused Turbo CTC and the diarization-focused 4.1 2B Plus. For developers, the release covers accurate transcription, fast on-device transcription and speaker-attributed transcripts.
TL;DR
- Size: 2B parameters, dense. LLM backbone uses YaRN extension to 16k positions. IBM recommends audio segments of up to about 2 minutes.
- Runs on: GPU via Transformers or vLLM, Apple Silicon via mlx-audio, local CPU/Mac via llama.cpp GGUF (Q8_0 recommended).
- Performance: #2 open-weight model on the Open ASR Leaderboard, 5.68% average English WER.
- Best: 0.99% WER on LibriSpeech Clean.
- Worst: 13.79% WER on the leaderboard’s private conversational set.
How does Granite Speech 5.0 2B work?
The model has 3 parts, per the model card. A speech encoder of 16 Conformer blocks is trained with CTC. It uses block attention over 4-second audio blocks.
A 2-layer window q-former then projects and downsamples the acoustic embeddings. Total downsampling is 10x, so the LLM sees embeddings at 10 Hz.
The LLM is an intermediate checkpoint of granite-4.0-1b-base. IBM extended it to 16k positions with YaRN and fine-tuned it with LoRA adapters.
Training used 390,000 hours of public audio plus synthetic data. The synthetic sets target Japanese ASR, keyword biasing and translation. Training took 40 days on 8 H100 GPUs.
What is new versus Granite Speech 4.1 2B?
- Language ID prompting: name the input language in the prompt to steer transcription.
- Keyword lists up to 2,000 terms: bias output toward product names, people and acronyms.
- Punctuation and capitalization on request: capitalization F1 runs from 90.16 to 99.51 across test sets.
- Robustness: better handling of noise, reverberation and distant microphones.
How does Granite Speech 5.0 2B perform on benchmarks?
On the Open ASR Leaderboard, IBM’s chart shows a 5.68% average English WER. That average covers 9 public and 2 private test sets, as of October 21, 2026. Only Qwen3-ASR-1.7B scores better, at 5.40%.
Note a source conflict. IBM’s blog reports 5.06% instead. That figure is unofficial, dated October 15, and averages public short-form sets only.
Per-set results range widely. The model scores 0.99% on LibriSpeech Clean and 2.21% on LibriSpeech Other. It scores 13.79% on the private conversational set.
On the noisy FFASR leaderboard, it ranks #4 of 42 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.
