SSyncDev

Granite Speech 5.0 2B

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.

Key specifications

Specification score
9.8
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted price stated.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Audio input
  • Text input

Produces

  • Text output

Capabilities

AudioOpen weights

What is Granite Speech 5.0 2B?

ollama run ibm/granite5-speech:2b audio.wav "transcribe the speech to text". That is the whole local setup for Granite Speech 5.0 2B, and it says something about where IBM aimed this release: a 2B-parameter speech-language model that runs on a laptop, a phone-class device via LiteRT, or a GPU server, all from one Apache 2.0 checkpoint.

IBM dates it 20 October 2026. Some coverage says the 21st.

Under the hood

A 16-block Conformer encoder (hidden size 1024, trained with CTC) feeds a two-layer windowed Q-Former that brings audio down to a 10 Hz embedding rate. The language model is an intermediate granite-4.0-1b-base checkpoint, extended with YaRN to 16K positions and fitted with LoRA adapters. Training used about 390,000 hours of public audio plus synthetic data.

It succeeds granite-speech-4.1-2b at the same size.

What it handles

  • Recognition in English, French, German, Spanish, Portuguese and Japanese.
  • Translation to and from English for those languages, plus English to Italian and English to Mandarin.
  • Keyword biasing with up to 2,000 terms, passed in the prompt: transcribe the speech to text. Keywords: <kw1>, <kw2>, ...

On language count, IBM's metadata says seven while the documented list has six for recognition. Test the language you need rather than trusting either tag.

Audio limits

Input is 16 kHz mono WAV. Keep segments to about two minutes and split longer recordings, for instance with voice-activity detection. There is no long-form mode.

Runtimes

Runtime Notes
Transformers 4.52.1+, AutoModelForSpeechSeq2Seq
vLLM vllm serve ibm-granite/granite-speech-5.0-2b --max-model-len 2048
llama.cpp GGUF, Q8_0 recommended
Ollama ibm/granite5-speech, 3.2 GB at Q8_0
MLX mlx-audio 0.5.9+ on Apple Silicon
LiteRT-LM 2.8 GB on-device bundle; 4,096-token context; audio encoder on CPU

An ONNX export has been reported, but IBM does not document it. Confirm before you plan an ONNX Runtime deployment.

Accuracy

IBM publishes leaderboard WER only as charts. In text, it gives punctuation and capitalisation gains over 4.1: punctuation error on LibriSpeech clean falls from 25.7 to 23.2, for example. Not much to compare against, which makes your own test essential.

Relation to TurboCTC

IBM's Granite Switch bundle embeds a different, smaller 470M TurboCTC speech encoder for transcription inside chat. This 2B model is the standalone speech-language model with translation and biasing. Do not confuse the two.

A quick bake-off

Take thirty two-minute clips per target language with the domain terms you care about. Run them through this model with and without a keyword list, and through Whisper Large v3 Turbo. Score WER overall and on the keywords specifically. Then time the Ollama build on the machine you will actually ship to.

More in automatic speech recognition and on IBM's provider page.

How to evaluate Granite Speech 5.0 2B

Compare the model on the work you need to complete.

Tasks and inputs

Granite Speech 5.0 2B is listed for Automatic speech recognition, Translation, Speech → Text. It accepts audio, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

English, French, German, Spanish, Portuguese and Japanese, with translation to and from English plus English to Italian and Mandarin.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

salesforce

SOV-Transcribe-1

SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.

— context

ElevenLabs

Scribe v3 Live

Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.

— context

NVIDIA

Nemotron 3.5 Transcribe 30B A4B

Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.

65.54K context

Mistral AI

Voxtral Mini Transcribe 3

Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.

65.54K context

Mistral AI

Voxtral Mini Transcribe Realtime 3

Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.

— context

Mistral AI

Voxtral Small Transcribe 3

Voxtral Small Transcribe 3 is Mistral's top-accuracy batch speech-to-text model for recorded audio, with context biasing and 50+ languages.

— context