granite-speech-5.0-2b

granite-speech-5.0-2b

Turns speech into text in six languages and translates speech to and from English, with keyword biasing and language prompting. A compact 2B speech-language model from IBM for enterprise audio.

Made by ibm-granite on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a compact speech-language model from IBM, called Granite Speech 5.0 2B. It transcribes speech in six languages and translates speech in both directions between English and the other five. IBM started from a middle checkpoint of granite-4.0-1b-base and aligned it to speech. Training used about 390,000 hours of audio taken from public sets, plus synthetic data.

Who it suits: It suits teams that process recorded or live speech in business settings and need transcripts, translations or keyword-aware output.

Industry | Who uses it | How they use it | Example
Enterprise operations | Teams that handle recorded calls and meetings | Transcribe speech into text | Turn a noisy call recording into a transcript with punctuation
Localization and media | Teams that work across several languages | Translate speech to and from English | Translate a spoken clip from Japanese into English text
Speech research | People who benchmark speech recognition | Compare error rates with the previous version | Check punctuation and capitalization scores on test sets

If you only need written text translated, or a language outside the supported set, you can skip it.

Where to use it:

• Python projects that use the transformers library
• Enterprise applications that process speech inputs
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Speech to text | Operations teams | A transcript with punctuation and capitalization
Speech translation | Localization teams | Text in the other language, in both directions with English
Keyword-aware transcription | Teams with product or name lists | Better capture of the terms on a list of up to 2000

How to set it up: What you need first:

• Python and the Hugging Face command line
• About 4.6 GB of memory for BF16 weights, or about 1.2 GB at 4-bit, which is an estimate
• Audio files or a stream to transcribe

1. Install the Hugging Face tools with pip.
2. Download the model files to your computer.
3. Load the model with transformers, following the usage section of the model card.

pip install -U huggingface_hub
hf download ibm-granite/granite-speech-5.0-2b

How to use it: You give the model audio and it returns text, either a transcript or a translation. The card says you can tell it the language of the audio with a language id prompt, and supply a keyword list of up to 2000 terms to bias recognition and translation.

Tool | What it does
Speech recognition | Turns audio in one of the six languages into text
Speech translation | Translates between English and the other languages, both ways
Language id prompting | Tells the model which language the audio is in
Keyword list biasing | Favors the terms on a list you provide

Example: You pass in a recording of an English earnings call and a short list of company and product names. The model returns a transcript with punctuation and capitals, and the list helps it get those names right.

For newcomers and experts: A good first try is one short English clip, transcribed as it is. Then try a clip in another supported language, and add a keyword list once the plain version works.

For experts, the model has 2.3B parameters and the card compares it with the previous 4.1 release at the same size. It reports higher accuracy on English short-form recognition, more resilience to noise and distant microphones, better long-form translation and a larger keyword list. For punctuation, PER counts inserted, deleted or substituted marks, and Cap-F1 scores capitalization on aligned word pairs.

Audio goes in
Optional language id and keyword list guide it
The model transcribes or translates
Text comes out with punctuation and capitals

Good to know:

Title: Punctuation error rate on LScln, lower is better
Previous 4.1 release: 25.70
This 5.0 release: 23.20

Title: Punctuation and capitalization against the previous release
Test set | PER 4.1 | PER 5.0 | Cap-F1 4.1 | Cap-F1 5.0
LScln | 25.70 | 23.20 | 89.71 | 90.16
LSoth | 22.27 | 21.22 | 91.26 | 91.65
VoxPopuli | 24.86 | 23.70 | 95.35 | 95.07
Earnings-22 | 22.87 | 21.64 | 95.19 | 94.98

• Scope: six languages for recognition, translation between English and each of them, and English as a source for Italian and Mandarin.
• Intended use: enterprise applications that process speech.
• Results: the card also shows charts for leaderboards, translation and keyword biasing, which are not reproduced here.
• License: Apache 2.0
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads files from Hugging Face, and the audio you give it is processed by your own code.
• Not for: languages outside the supported set.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 2.3B
• Weights: about 4.6 GB in BF16; about 1.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 10,081
• Likes: 84
• License: apache-2.0
• Languages: en, fr, de, es, pt, ja
• Based on: ibm-granite/granite-4.0-1b-base
• Quantized versions: 9
• Fine-tunes: 1
• Tags: transformers, safetensors, granite_speech, automatic-speech-recognition, multilingual

Links

• Model: https://huggingface.co/ibm-granite/granite-speech-5.0-2b
• Files: https://huggingface.co/ibm-granite/granite-speech-5.0-2b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ibm-granite/granite-speech-5.0-2b

Then follow the usage in its model card. Nothing in this kit runs by itself.