IBM Releases Granite 4.0 1B Speech, a Compact Multilingual Model for Edge Devices
IBM has unveiled Granite 4.0 1B Speech, a compact automatic speech recognition model that achieves top-tier performance while running on resource-constrained devices. The model, released on March 9, 2026, supports six languages and ranks #1 on the OpenASR leaderboard despite using significantly fewer parameters than competing solutions.
The 1-billion-parameter model represents IBM's push into edge deployment for enterprise speech applications. It supports English, French, German, Spanish, Portuguese, and Japanese, with new Japanese ASR capabilities and keyword list biasing for improved recognition of names and acronyms.
Performance Benchmarks
Granite 4.0 1B Speech achieves its top ranking on OpenASR using Word Error Rate (WER) measurements, where lower scores indicate better accuracy. The model delivers higher English transcription accuracy and faster inference through speculative decoding compared to its predecessor, granite-speech-3.3-2b.
The architecture uses a speech encoder with 16 conformer blocks trained with Connectionist Temporal Classification on character-level targets. It includes block-attention processing on 4-second audio blocks and self-conditioned CTC from the middle layer.
Deployment and Licensing
Released under the Apache 2.0 license, the model offers native support in Hugging Face Transformers and vLLM, making it suitable for edge, on-device, and latency-sensitive enterprise use cases. This positions IBM against larger models from Google, OpenAI, and Anthropic in the speech recognition space while targeting a different market segment focused on efficient local deployment.
The model is available on Hugging Face at ibm-granite/granite-4.0-1b-speech.