IBM Releases Granite Speech 5.0 2B for Transcription

This 2B-parameter model from IBM converts speech into text across 6 languages and is distributed under the Apache 2.0 license.

The release also includes updates to two existing models: the Turbo CTC for edge computing and the 4.1 2B Plus for diarization.

Granite Speech 5.0 2B provides developers with capabilities for accurate transcription, fast on-device transcription, and speaker-attributed transcripts.

The model utilizes a 2B parameter count and an LLM backbone with a YaRN extension to 16k positions, with IBM recommending audio segments of up to approximately 2 minutes.

It can run on GPUs via Transformers or vLLM, Apple Silicon via mlx-audio, and locally on CPU/Mac via llama.cpp GGUF, with Q8_0 recommended.

The model achieved the #2 open-weight model ranking on the Open ASR Leaderboard with a 5.68% average English Word Error Rate (WER), and a best of 0.99% WER on LibriSpeech Clean.

Granite Speech 5.0 2B features a speech encoder with 16 Conformer blocks trained with CTC, using block attention over 4-second audio blocks.

A 2-layer window q-former projects and downsamples acoustic embeddings, resulting in a total downsampling of 10x, allowing the LLM to process embeddings at 10 Hz.

The LLM component is an intermediate checkpoint of granite-4.0-1b-base, extended to 16k positions with YaRN and fine-tuned using LoRA adapters.

Training involved 390,000 hours of public audio data and synthetic data targeting Japanese ASR, keyword biasing, and translation, completed in 40 days on 8 H100 GPUs.

New features compared to Granite Speech 4.1 2B include language ID prompting to steer transcription by specifying the input language, and keyword lists of up to 2,000 terms to bias output.