Overview
IBM Granite Speech 5.0 2B is a speech-language model IBM open-sourced in October 2026, with 2B parameters under the Apache 2.0 licence and weights published on Hugging Face. The architecture pairs a 16-layer conformer speech encoder and a 2-layer q-former projector with granite-4.0-1b-base as the language backbone, with YaRN extending positions to 16k.
It handles speech recognition in English, French, German, Spanish, Portuguese and Japanese, bidirectional speech translation among those languages, and additional English-to-Italian and English-to-Mandarin translation. Compared with the previous generation it improves short-utterance accuracy, holds up better against noise, reverb and far-field microphones, adds an input language hint, and raises keyword biasing to 2,000 entries.
Training used 390,000 hours of audio across 40 days on eight H100 GPUs in IBM Blue Vela. Inference runs through Transformers, vLLM, the llama.cpp GGUF build or mlx-audio on Apple Silicon, and audio segments are best kept under about two minutes before decoding.
Key Features
- Six-language transcription: English, French, German, Spanish, Portuguese and Japanese, covering meetings, support calls and podcasts
- Bidirectional speech translation: Those languages translate in both directions, with English to Italian and English to Mandarin on top
- 2B open weights: Apache 2.0 licensed, with weights and usage notes published directly on Hugging Face
- Keyword biasing: Up to 2,000 keyword hints, which helps with product names and domain terminology
- Input language hint: Multilingual audio can declare its input language, cutting misdetection within a clip
- Several inference backends: Examples ship for Transformers, vLLM, the llama.cpp GGUF build and mlx-audio, covering CPU and Apple Silicon
Use Cases
- Multilingual transcription of meetings and interviews
- Speech translation for cross-border support
- Subtitles for podcasts and video
- Localisation into Japanese and European languages
- On-device or private transcription
- Handling sensitive recordings inside a local network
Pros
- The Apache 2.0 licence leaves commercial use and retraining unrestricted
- At 2B parameters it runs on a single consumer GPU or even a CPU
- Six-language transcription plus bidirectional translation covers cross-border work in one model
- Keyword biasing up to 2,000 entries leaves room for industry vocabulary
- IBM documents four ways to run it: Transformers, vLLM, llama.cpp and mlx-audio
- Long recordings split at two minutes and decode in batches
Pricing
The weights are free on Hugging Face under the Apache 2.0 licence. Running costs depend on your own hardware and deployment, with no usage-based fees.
Summary
Granite Speech 5.0 2B squeezes multilingual transcription and translation into 2B parameters, and the Apache 2.0 licence keeps commercial use uncomplicated. It is best read as a component you can run locally: meeting transcription, cross-border support and subtitles all fit on one GPU or a CPU, and sensitive recordings never have to leave the local network.
Version History
- Granite Speech 5.0 2B open-sourced for six-language speech (2026-10-20): IBM open-sourced Granite Speech 5.0 2B in October 2026 under the Apache 2.0 licence, with weights on Hugging Face. It transcribes English, French, German, Spanish, Portuguese and Japanese and translates bidirectionally among them, plus English to Italian and English to Mandarin. Training used 390,000 hours of audio over 40 days on eight H100 GPUs. Keyword biasing reaches 2,000 entries, and it runs through Transformers, vLLM, llama.cpp and mlx-audio.