Development ·
IBM releases Granite Speech 5.0 2B with multilingual keyword guidance
The Apache 2.0 speech model adds language prompting and supports keyword lists of up to 2,000 terms for transcription and speech translation.
Why it matters
Developers can evaluate an open model with vocabulary controls for domain-specific audio, including deployments where they choose the inference infrastructure.
What to watch
Independent tests on accented, noisy and specialist recordings, plus hardware measurements and errors introduced by keyword biasing.
Still uncertain. Accuracy, latency and memory use were not independently measured. The model-card release date is October 20; the broader announcement appeared October 21.
Impact assessment reviewed 2026-10-22.
- Who it helps
- Developers building transcription, translation or speech-processing applications.
- Availability
- Apache 2.0 weights are public. The model card dates release October 20, 2026; the family announcement followed October 21. Local compute or hosted inference costs still apply.
Control the transcription task
The card supports English, French, German, Spanish, Portuguese and Japanese. You can specify the input language and supply up to 2,000 keywords. For a meeting with unusual project names, that makes a controlled vocabulary a concrete input to evaluate. A keyword list can bias the result; it does not guarantee the words were spoken.
Check audio length and runtime
The documentation recommends splitting audio into segments of roughly two minutes. It lists Transformers 4.52.1 or later, vLLM, llama.cpp and mlx-audio; the Apple Silicon example requires mlx-audio 0.5.9 or later plus jinja2. IBM recommends Q8_0 for its quantized variant and reports accuracy loss with lower-bit versions on longer audio. These are documented constraints, not a measured hardware recommendation.
Granite Speech 5.0 2B model card ↗Granite Speech 5 family announcement ↗
What you can do
- Compare a short recording with a human-checked transcript before assessing accuracy, memory use or cost. Open reference ↗
Where the evidence stops
Model card and announcement checked; no transcription, translation or hardware benchmark was executed. Language support and author benchmarks do not establish accuracy for a particular speaker or profession.
Sources read. Sources checked October 22, 2026 UTC. No external software workflow, model inference or reported experiment was executed.
Sources & further reading
Hugging Face · official
Granite Speech 5.0 2B model card ↗
Native model card establishes the October 20 release, Apache 2.0 license, language coverage and runtime constraints. No weights were downloaded or run.
IBM Granite · page checked 2026-10-22 · Published 2026-10-20
Hugging Face · official
Granite Speech 5 family announcement ↗
Original announcement covers three distinct models; this story identifies the new 5.0 2B model separately from the other updates.
IBM Granite · page checked 2026-10-22 · Published 2026-10-21
More sources reviewed (1)
Hugging Face · analysis
IBM Granite creator index ↗
Creator activity followed one level beyond the demo to original Switch and Speech announcements.
IBM Granite · page checked 2026-10-22
Coverage note: The linked primary evidence was inspected. Independent execution and a complete cross-platform discussion sample are missing; no original X source was verified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
2026-10-20: The Apache 2.0 speech model adds language prompting and supports keyword lists of up to 2,000 terms for transcription and speech translation.
Assessment: Developers can evaluate an open model with vocabulary controls for domain-specific audio, including deployments where they choose the inference infrastructure.