IBM has open-sourced Granite Speech 5.0 2B, a 2-billion-parameter speech-to-text and speech translation model released under Apache 2.0 on October 21. It averages 5.06% word error rate on the Open ASR Leaderboard's public English test sets, second among open-weight models behind Qwen3-ASR-1.7B. IBM also updated two companion models: a faster, now multilingual 470M Turbo CTC model and a meeting-focused 2B Plus model.

  • Model: granite-speech-5.0-2b, 2B parameters, built on IBM's granite-4.0-1b-base language model

  • Accuracy: 5.06% average English WER on the leaderboard's 9 public short-form sets

  • Languages: transcription in English, French, German, Spanish, Portuguese and Japanese; translation to and from English, plus English to Italian and Mandarin

  • License: Apache 2.0, weights on Hugging Face

  • Runs on: Transformers, vLLM, llama.cpp (GGUF) and mlx-audio for Apple Silicon

What is Granite Speech 5.0?

Granite Speech is IBM's family of open speech recognition models. The name first appeared in August with a 470M-parameter Turbo CTC model built for raw speed, covered in the Voice AI Space week 35 roundup. The October release adds the larger model, granite-speech-5.0-2b, the successor to granite-speech-4.1-2b at the same parameter count.

It is a speech-language model: an audio encoder aligned to a small Granite language model, prompted with plain-text instructions. IBM says it was trained on 390,000 hours of public audio plus synthetic data for Japanese, keyword biasing and translation.

What changed from Granite Speech 4.1

  • Accuracy and robustness: higher accuracy on English short-form benchmarks, and better handling of noise, reverberation and distant microphones, according to IBM.

  • Keyword biasing up to 2,000 terms: product names, people and acronyms are passed in the prompt, in every supported language.

  • Language ID prompting: naming the input language in the prompt steers multilingual transcription.

  • Punctuation and capitalization on request: IBM reports capitalization F1 from 89.7 to 99.5 across its test sets.

  • Longer translation: improved long-form and conversational speech translation.

How does Granite Speech 5.0 compare with other open speech-to-text models?

Model

Parameters

Weights license

Avg English WER (9 public sets)

RTFx

Qwen3-ASR-1.7B (Alibaba Qwen)

1.7B

Apache 2.0

4.94%

819

Granite Speech 5.0 2B (IBM)

2B

Apache 2.0

5.06%

545

Canary-Qwen-2.5B (NVIDIA)

2.5B

CC BY 4.0

5.38%

862

Parakeet TDT 0.6B v3 (NVIDIA)

0.6B

CC BY 4.0

5.58%

6,132

Granite Speech 5.0 470M Turbo CTC (IBM)

470M

Apache 2.0

5.78%

21,174

Whisper large-v3 (OpenAI)

1.55B

Apache 2.0

6.66%

470

All figures from the Hugging Face Open ASR Leaderboard, 21 October 2026 version, same test sets and setup for every model. WER is averaged over the 9 public English short-form sets, the method behind IBM's 5.06%. Lower WER is better; RTFx is seconds of audio processed per second of compute, higher is faster. Licenses as listed on each model's Hugging Face page.

The leaderboard's default view also averages in two private test sets and ranks closed APIs alongside open models. On that view Granite Speech 5.0 2B scores 5.68% and places 10th, behind commercial services from Otter.ai, Zoom, Microsoft, ElevenLabs, AssemblyAI and others. Among models with downloadable weights, it is second on both views.

The trade-off between the two IBM models is speed. The 2B model is the more accurate of the pair; Turbo CTC processes audio roughly 39 times faster on the same benchmark. IBM reports the revised Turbo CTC at 19,870 RTFx, about 1.6 times its August launch speed, and says it now transcribes seven languages with native punctuation. A separate non-commercial variant of Turbo CTC (CC BY-NC-SA 4.0) scores 5.44% on the same sets.

Granite Speech 4.1 2B Plus for meetings and calls

The third model, granite-speech-4.1-2b-plus, is a drop-in weight update for multi-speaker recordings. IBM's figures against the previous release: word diarization error rate down 7% (4.40 to 4.10), WER on long-form test sets down 24% (13.53 to 10.22), errors on biased keywords down 45% with 20-term lists, and average timestamp shift cut by more than half.

Where can you use it?

All three models are on Hugging Face under Apache 2.0, with native support in Transformers. The 2B model also runs on vLLM with an OpenAI-compatible server, on llama.cpp through a GGUF build (IBM recommends Q8_0, noting that lower-bit quantizations lose accuracy on audio longer than about a minute), and on Apple Silicon through mlx-audio. Hugging Face hosts a scale-to-zero Inference Endpoint for testing. Turbo CTC has an in-browser WebGPU streaming demo.

What it means for teams building voice agents

Apache 2.0 weights allow commercial self-hosting without usage fees, which matters for teams that keep audio on their own infrastructure. The 2B model is built for batch and near-real-time transcription rather than streaming: the model card recommends audio segments of up to about two minutes, with longer recordings split first. For a refresher on where speech-to-text sits in a voice agent pipeline, see How Voice Agents Work: STT, LLM, TTS, WebRTC.

What's missing

  • No streaming latency figures for the 2B model.

  • No managed IBM API or per-minute price in the announcement; the models are weights only.

  • The leaderboard ranking covers English. Multilingual and translation results come from IBM's own evaluations on CommonVoice, CoVoST and FLEURS.

Sources