Granite 5.0 Speech

Granite-Speech-5.0-2B is a compact and efficient speech-language model from IBM, purpose-built for multilingual automatic speech recognition (ASR) and bidirectional automatic speech translation (AST) for English, French, German, Spanish, Portuguese, and Japanese. It was created by modality-aligning an intermediate checkpoint of granite-4.0-1b-base to speech, and trained on 390,000 hours of audio from public corpora, plus synthetic datasets tailored to support Japanese ASR, keyword-biased ASR, and speech translation.

Compared to its predecessor, granite-speech-4.1-2b, this model has the same parameter count and adds the following improvements:

• Higher transcription accuracy for English short-form ASR, and increased robustness to noise, reverberation, and distant-microphone speech.
• Input language ID prompting for multilingual ASR.
• Improved long-form and conversational speech translation.
• Larger keyword lists (up to 2000) for multilingual keyword-biased ASR and AST.

Running

Pass a path to a .wav file and a task prompt directly on the command line. To run the default quantization:

ollama run ibm/granite5-speech:2b /path/to/audio.wav "transcribe the speech with proper punctuation and capitalization."

To run a specific quantization (e.g., Q8_0):

ollama run ibm/granite5-speech:2b-q8_0 /path/to/audio.wav "transcribe the speech with proper punctuation and capitalization."

The model accepts mono, 16 kHz .wav audio. To trigger speech processing, include the <|audio|> token in your prompt. If the model receives an unfamiliar or malformed prompt, it falls back to transcription by default. For best results, transcribe audio in segments of up to ~2 minutes; split longer recordings (e.g., with voice activity detection) before decoding.

Preferred Prompts by Task

For ASR with language ID, <language> is one of English, French, German, Spanish, Portuguese, or Japanese. For AST, <language> is one of English, French, German, Spanish, Japanese, Italian, or Mandarin.

Prompts for raw ASR, punctuated ASR, language ID, and keyword-biased ASR/AST can also be written in other supported languages (e.g., Transcrivez la parole avec une ponctuation et des majuscules appropriées.). The punctuated AST prompt supports English only.

Supported Languages

• ASR: English, French, German, Spanish, Portuguese, and Japanese.
• AST: speech translation to and from English for the same languages, plus English-to-Italian and English-to-Mandarin.

Intended Use

Granite-Speech-5.0-2B is designed for enterprise applications that process speech inputs, in particular speech-to-text and speech translation for English, French, German, Spanish, Portuguese, and Japanese.

Capabilities

• Multilingual ASR — High-accuracy transcription across six languages, powered by a conformer encoder trained with CTC (character and BPE output heads) and a window query transformer (q-former) speech projector.
• Punctuation and Capitalization — Optional punctuated, written-form transcripts via prompt.
• Language ID Prompting — Name the spoken language in the prompt to improve multilingual ASR.
• Bidirectional Speech Translation (AST) — Translate speech to and from English, with improved long-form and conversational performance.
• Keyword Biasing — Improved recognition of names, acronyms, and technical jargon when supplied with a keyword list of up to 2000 entries, for both ASR and AST.
• Noise Robustness — Improved accuracy on noisy, reverberant, and distant-microphone speech.

Performance

Punctuation and capitalization — PER is Punctuation Error Rate (%, lower is better). Cap-F1 is capitalization F1 (%, higher is better). Best results in each pair are in bold.

Metrics are as defined in LibriSpeech-PC. Cap-F1 is computed on Levenshtein-aligned matching word pairs rather than fully matching sentences, so evaluation is possible even in the presence of ASR errors. The CV-DE Cap-F1 of 99.51 is on German, where noun capitalization is required.

For English ASR results on the Open ASR leaderboard, noisy and reverberant speech results on the FFASR leaderboard, multilingual ASR on CommonVoice, speech translation (CoVoST X→En, FLEURS En→X), and keyword-biasing F1 comparisons, see the model card.

Safety

The model’s design limits how audio inputs can influence the system. If an unfamiliar or malformed prompt is received, the model ignores it and performs transcription, which is the default fallback mode. IBM recommends using this model for ASR and translation tasks, and using it alongside Granite Guardian to detect and flag risks in prompts and responses.

Learn more

• Developers: IBM Granite Speech Team
• Website: Granite Docs
• Blog: Granite Speech 5
• HuggingFace: ibm-granite/granite-speech-5.0-2b
• Release Date: October 20, 2026
• License: Apache 2.0