What is Granite Speech 5.0 2B?
Granite Speech 5.0 2B is a speech-language model from IBM, released on 20 October 2026 under Apache 2.0, that transcribes English, French, German, Spanish, Portuguese and Japanese and translates speech to and from English for those languages, plus English into Italian and Mandarin. It replaces Granite Speech 4.1 2B at the same size and was trained on 390,000 hours of audio from public corpora plus synthetic data.
It is built from a 16-layer Conformer speech encoder, a small adapter and a Granite 4.0 1B language model, and is driven by text prompts: one prompt asks for raw text, another for punctuation and capitals, another names the input language, and a keyword list of up to 2,000 terms biases it towards names and jargon. It runs in Transformers, vLLM, llama.cpp, Ollama as ibm/granite5-speech (a 3.2 GB Q8_0 build), mlx-audio on Apple Silicon and LiteRT-LM on Android.
Is Granite Speech 5.0 2B worth it?
Granite Speech 5.0 is accurate on English and unusually good at the words generic models get wrong. IBM reports a 5.06% average English WER on the Open ASR Leaderboard's public short-form sets as of 21 October 2026, second among open-weight models, at 570 times real time; its predecessor scored 5.33%. IBM's chart puts Qwen3-ASR-1.7B first at 4.94%, so this is a close second rather than a new leader, and the figures are IBM's own runs.
The distinctive feature is keyword biasing. You append a list of up to 2,000 terms to the prompt, in any of its languages, and the model leans towards them: product names, drug names, colleague surnames. IBM's earlier 4.1 Plus update cut WER on biased words from 11.79% to 6.48% with 20-term lists, and 5.0 raises the list size.
Translation is the second reason to choose it. One checkpoint transcribes and translates, so a German meeting can come out as English text in a single pass. IBM also reports better robustness to noise, echo and distant microphones; on the noisy-speech FFASR board IBM's chart ranks it fourth of 42 at 14.80% WER.
The catch: it is short-form at heart. IBM advises segments of about 2 minutes, with longer recordings split by a voice detector first, and lower-bit GGUF quantisations lose accuracy beyond about a minute, so Q8_0 is the floor. Six transcription languages is narrow next to Qwen3-ASR's 30, and the 2B card offers no timestamps or speaker labels; IBM puts those in the separate Granite Speech 4.1 2B Plus.
Our take: Granite Speech 5.0 suits enterprise English and European work where a glossary can be supplied, such as medical, legal or support calls; for wide language coverage or long recordings in one go, other open models fit better.
Where it earns its place
A clinical-documentation team at a hospital group
Transcribe doctor dictation with drug and procedure names spelt correctly
Pass the formulary as a keyword list of up to 2,000 terms with each 2-minute segment. Running it on-premise under Apache 2.0 keeps patient audio inside the hospital network.
An operations manager at a Franco-German manufacturer
Get English minutes from shop-floor meetings held in French and German
Use the translate-to-English prompt so one pass produces English text; split recordings into 2-minute chunks with a voice detector first, as IBM advises.
A mobile developer building an offline note-taking app
Run transcription on an Android phone without a network
Ship IBM's LiteRT-LM conversion, a single 2.8 GB file with an int8 language model, whose context holds 4,096 tokens, about 2 minutes of audio plus the reply.
Before you commit
- Setup
- pip install transformers torchaudio soundfile and load ibm-granite/granite-speech-5.0-2b, or run ollama run ibm/granite5-speech:2b audio.wav with a prompt. Audio must be 16 kHz mono.
- Integrations
- Transformers 4.52.1+, vLLM with an OpenAI-compatible server, llama.cpp llama-server with /v1/audio/transcriptions, Ollama, mlx-audio 0.5.9+, LiteRT-LM 0.19.1; IBM suggests pairing it with Granite Guardian in production.
- Data and privacy
- Apache 2.0 with training data drawn from public corpora and IBM-generated synthetic sets. Self-hosted, audio stays on your hardware; the Hugging Face endpoint follows Hugging Face's terms.
- Cost at scale
- Free weights; the Q8_0 build is 3.2 GB, so it fits small GPUs and Apple laptops. At IBM's measured 570 times real time, a million minutes is about 29 GPU-hours of compute before overheads.
- Limits
- Six transcription languages; translation into Italian and Mandarin from English only; best under about 2 minutes per segment; 4,096-token context in the LiteRT build; no diarization or timestamps.
What’s good
- 5.06% average English WER, second among open models in IBM's October chart
- Keyword biasing with lists of up to 2,000 terms
- Transcribes and translates in one model
- Apache 2.0; runs in Ollama, llama.cpp, mlx-audio and on Android
What’s not
- Six transcription languages
- Best with segments of about 2 minutes; long audio needs splitting
- Quantisations below Q8_0 lose accuracy on clips over a minute
- No timestamps or speaker labels in this model
- Benchmarks are IBM's own
Best for
Enterprises transcribing English and major European languages with domain vocabulary they can list in advance.
Granite Speech 5.0 2B alternatives
Granite Speech 5.0 2B: common questions
- Is Granite Speech 5.0 2B worth it?
- Granite Speech 5.0 is the open transcriber to pick when names and jargon matter more than language count. We rate it 4.0 out of 5.
- How much does Granite Speech 5.0 2B cost?
- Free under Apache 2.0. A Hugging Face Inference Endpoint is available that scales to zero when idle; its hourly cost depends on the hardware chosen.
- What are the downsides of Granite Speech 5.0 2B?
- Six transcription languages; best with segments of about 2 minutes; long audio needs splitting; quantisations below Q8_0 lose accuracy on clips over a minute; no timestamps or speaker labels in this model; benchmarks are IBM's own.
- Who is Granite Speech 5.0 2B best for?
- Enterprises transcribing English and major European languages with domain vocabulary they can list in advance.
- How much does Granite Speech 5.0 2B cost for a team or in production?
- Free weights; the Q8_0 build is 3.2 GB, so it fits small GPUs and Apple laptops. At IBM's measured 570 times real time, a million minutes is about 29 GPU-hours of compute before overheads.
- Is Granite Speech 5.0 2B safe for company data?
- Apache 2.0 with training data drawn from public corpora and IBM-generated synthetic sets. Self-hosted, audio stays on your hardware; the Hugging Face endpoint follows Hugging Face's terms.
- What are the best Granite Speech 5.0 2B alternatives?
- The closest alternatives we have reviewed are Qwen3-ASR (1.7B and 0.6B), Cohere Transcribe (03-2026), Parakeet TDT 0.6B v3, Whisper large-v3-turbo and ElevenLabs Scribe v2.
Answers checked Nov 7, 2026.
Make Granite Speech 5.0 2B? Show this rating on your site
