IBM Releases Granite Speech 5.0 2B: An Open Multilingual Speech-to-Text Model for Transcription and Translation
IBM has released Granite Speech 5.0 2B, a new open speech-language model for transcription and translation. The 2B-parameter model converts speech into text across six languages and ships under the Apache 2.0 license. Alongside the new release, IBM updated two existing models: the edge-focused Turbo CTC and the diarization-focused 4.1 2B Plus. Together, the release targets three developer needs: accurate transcription, fast on-device transcription, and speaker-attributed transcripts.
TL;DR
- Size: 2B parameters, dense. The LLM backbone uses YaRN extension to 16k positions. IBM recommends audio segments of up to about two minutes.
- Runs on: GPU via Transformers or vLLM, Apple Silicon via mlx-audio, and local CPU/Mac via llama.cpp GGUF (Q8_0 recommended).
- Performance: Ranked #2 among open-weight models on the Open ASR Leaderboard, with 5.68% average English WER.
- Best result: 0.99% WER on LibriSpeech Clean.
- Worst result: 13.79% WER on the leaderboard’s private conversational set.
How Does Granite Speech 5.0 2B Work?
The model pairs a speech encoder with a dense 2B-parameter language-model backbone. The backbone uses YaRN (Yet another RoPE extensioN) to extend positional context to 16k tokens, which supports longer audio inputs. For practical use, IBM recommends keeping audio segments to roughly two minutes or less to stay within reliable performance bounds. The architecture is designed for both transcription and translation tasks, making it a single model for multilingual speech workflows.
Deployment Options
Granite Speech 5.0 2B is designed to run across a range of hardware configurations:
- GPU: Transformers or vLLM for high-throughput server-side inference.
- Apple Silicon: mlx-audio for native Mac inference.
- Local CPU/Mac: llama.cpp with GGUF quantization, with Q8_0 recommended for the best balance of quality and speed.
This flexibility supports both cloud deployments and privacy-sensitive on-device transcription.
Performance Highlights
On the Open ASR Leaderboard, Granite Speech 5.0 2B ranks second among open-weight models, with an average English word error rate (WER) of 5.68%. Its strongest result is 0.99% WER on LibriSpeech Clean, reflecting near-perfect transcription on clean read speech. Its weakest result is 13.79% WER on the leaderboard’s private conversational set, which reflects the difficulty of spontaneous, noisy, multi-speaker conversation.
Related Model Updates
IBM also refreshed two companion models:
- Turbo CTC: An edge-focused model optimized for fast, low-resource transcription.
- 4.1 2B Plus: A diarization-focused model for speaker-attributed transcripts.
Why It Matters
The release expands IBM’s open Granite family into multilingual speech with a permissive Apache 2.0 license. For developers, it offers a compact 2B-parameter option that covers accurate transcription, on-device inference, and speaker-attributed output, with deployment paths spanning GPUs, Apple Silicon, and local CPUs.
via MarkTechPost