WORLDTECHIBM releases Granite Speech 5.0 2B, an open transcription and translation model
AI illustration
IBM has released Granite Speech 5.0 2B, an open 2-billion-parameter model that turns speech into text in six languages, according to a report by one outlet. The Armonk-based software and information technology consulting company is positioning the model for transcription and translation work.
Release scope and benchmark standing
The model is distributed under the Apache 2.0 licence and is presented as an open speech recognition and translation system. IBM's chart on the Open ASR Leaderboard places it second, with an average English word error rate of 5.68%. On the leaderboard's private conversational set, its worst result is 13.79% WER.
IBM has also updated two existing models: Turbo CTC, which targets edge devices, and 4.1 2B Plus, which focuses on diarization. For developers, the release is aimed at accurate transcription, fast on-device transcription and speaker-attributed transcripts.
Architecture and training
According to the model card, the system has three parts. A speech encoder of 16 Conformer blocks is trained with CTC and applies block attention over 4-second audio blocks. A two-layer window Q-former then projects and downsamples the acoustic embeddings, a tenfold reduction that lets the language model see embeddings at 10 Hz.
The language model is an intermediate checkpoint of granite-4.0-1b-base, which IBM extended to 16k positions with YaRN and fine-tuned with LoRA adapters. Training used 390,000 hours of public audio plus synthetic data targeting Japanese ASR, keyword biasing and translation, and took 40 days on 8 H100 GPUs.
Features and deployment options
Keyword lists of up to 2,000 terms can bias output toward product names, people and acronyms, and punctuation and capitalization are available on request. IBM recommends audio segments of up to about 2 minutes. The model runs on GPU through Transformers or vLLM, on Apple Silicon through mlx-audio, and locally through llama.cpp GGUF, with Q8_0 recommended.
IBM's figures show capitalization F1 between 90.16 and 99.51 across test sets.
Open questions
- How does the model perform on languages beyond the six it covers?
- Will independent benchmarks confirm IBM's leaderboard figures?