IBM's Granite Speech 5.0 2B Transcribes Six Languages, Translates Speech and Takes 2,000-Word Keyword Lists

IBM has released Granite Speech 5.0 2B, an open model that turns speech into text and translates it, rather than generating voices. It does automatic speech recognition in English, French, German, Spanish, Portuguese and Japanese, and translates speech between English and each of those languages, plus from English into Italian and Mandarin. The weights are on Hugging Face under the Apache 2.0 licence. IBM’s announcement went out on October 21; the model card gives October 20 as the release date.

The model replaces Granite Speech 4.1 2B at the same size. IBM’s pitch is enterprise transcription: noisy rooms, distant microphones, long recordings and vocabulary full of product names and acronyms.

What changes from the 4.1 model

IBM lists five improvements over its predecessor:

• Higher accuracy on English short-form tests, and better handling of noise, reverberation and distant-microphone speech.
• Language-ID prompting: name the input language in the prompt to steer multilingual transcription.
• Translation that holds up over longer, more conversational speech, not just short read sentences.
• Keyword lists of up to 2,000 terms to bias transcription and translation toward names, products and acronyms in every supported language.
• Punctuation and capitalization on request. IBM reports capitalization F1 scores between 89.7 and 99.5 across its test sets.

Prompts are plain instructions such as “transcribe the speech to text” or “translate the speech to” a named language, with an optional keyword list appended. The card says the model falls back to transcription if it receives an unfamiliar or malformed prompt.

IBM’s accuracy claims

IBM says the model averages a 5.06% word error rate on the public English short-form test sets of the Open ASR Leaderboard, which it says places it second among open-weight speech recognition models. That ranking is IBM’s own reading of the leaderboard as of October 21. The model card also charts multilingual results on Common Voice, speech translation scores on CoVoST and FLEURS, and noisy-speech results on the FFASR leaderboard, all published by IBM as images rather than tables.

Inside the model

Granite Speech 5.0 2B has three parts. A speech encoder of 16 Conformer blocks, trained with connectionist temporal classification, turns audio into acoustic embeddings. A two-layer query transformer then compresses those embeddings to 10 per second of audio. Finally, an intermediate checkpoint of IBM’s Granite 4.0 1B base language model, extended to 16,000 positions and adapted with LoRA, writes the transcript or translation.

IBM trained it on 390,000 hours of audio. Most of that is public corpora such as Common Voice, Multilingual LibriSpeech, LibriHeavy, VoxPopuli, YODAS and the Granary sets. The rest is synthetic data built for Japanese recognition, keyword biasing and translation. Some translation targets were produced with other models, including Phi-4 and MADLAD. The card says training took 40 days on eight Nvidia H100 GPUs.

Two-minute segments and 8-bit quantization

The model runs natively in Hugging Face Transformers and is also supported in vLLM, llama.cpp and mlx-audio on Apple Silicon. IBM recommends transcribing audio in segments of up to about two minutes and splitting longer recordings first. For llama.cpp users, it recommends the 8-bit quantization, saying lower-bit versions lose accuracy on audio longer than about a minute.

Two companion updates

The same announcement refreshed two other Granite Speech models. Granite Speech 5.0 470M TurboCTC, a small encoder built for fast and on-device transcription, now handles seven languages and adds punctuation and capitalization. IBM says it now runs at 19,870 RTFx on the Open ASR Leaderboard, up from about 12,600 at launch in August.

Granite Speech 4.1 2B Plus, aimed at meetings and calls with several speakers, gets new weights with the same interface. In IBM’s tests, its word diarization error rate drops 7%, long-form word error rate drops 24% and timestamp drift is cut by more than half.

IBM’s advice for choosing between them is simple: the 5.0 2B model for accuracy, translation and keyword biasing; TurboCTC for raw speed or devices; and 4.1 2B Plus when knowing who spoke matters. All three are Apache 2.0, and IBM suggests pairing them with its Granite Guardian models in production.