Granite Speech 5.0: Specs, License & How to Run

Granite Speech 5.0 is IBM’s October 2026 speech family. Official hosts we opened on 2026-10-24: ibm-granite/granite-speech-5.0-2b (created 2026-10-07T16:10:49Z; card Release Date: October 20, 2026; 48 likes / 1,125 downloads), the speaker-attribution sibling ibm-granite/granite-speech-4.1-2b-plus (101 likes / 34,258 downloads; last modified 2026-10-21), the edge CTC card ibm-granite/granite-speech-5.0-470m-turboctc (100 likes / 58,154 downloads), and the blog Granite Speech 5.0.

This page hangs off Otter.ai alternatives and Descript alternatives. That mapping is editorial. Otter is a hosted meeting agent. Descript is a timeline editor. These cards transcribe, translate, or tag speakers on an audio file. We do not claim product parity. Related local STT: whisper.cpp. Related diarization-only pipeline: pyannote community-1. Same-site file ASR peer: Nemotron 3.5 Transcribe — Nemotron 3.5 Transcribe vs Granite Speech 5.0.

Family split (do not mix the ids)

Card	Job	Likes / downloads (2026-10-24)	Params (API)
granite-speech-5.0-2b	Multilingual ASR + bidirectional AST	48 / 1,125	2,313,207,132
granite-speech-4.1-2b-plus	Speaker-attributed ASR ([Speaker N]:) + timestamps	101 / 34,258	~2.11B
granite-speech-5.0-470m-turboctc	Fast multilingual CTC ASR	100 / 58,154	~473M

5.0-2B (card): English, French, German, Spanish, Portuguese, Japanese ASR; translation to and from English in those languages, plus English-to-Italian and English-to-Mandarin. Trained on 390,000 hours of public audio plus synthetic data. Language-id prompting, keyword lists up to 2,000 terms, optional punctuation/capitalization.

4.1-2B Plus is the speaker-attribution checkpoint. It keeps the 4.1 name because the architecture and prompt format stay drop-in. The family blog (20 October 2026) quotes vendor deltas versus the previous Plus revision: WDER 4.40 → 4.10, cpWER 30.50 → 28.48, long-form WER 13.53 → 10.22. Those rows stay self-reported.

What we verified (2026-10-24)

Field	5.0-2B	Source
Host	huggingface.co/ibm-granite/granite-speech-5.0-2b (not gated)	HF API
Created / last modified	2026-10-07T16:10:49Z / 2026-10-23T11:27:40Z	same
Card release date	October 20, 2026	README
License tag	apache-2.0	HF cardData
Pipeline tag	automatic-speech-recognition	HF API
usedStorage	5,040,872,787 bytes (~4.7 GiB)	HF API
Base	granite-4.0-1b-base (card)	card

How to run (from the 5.0-2B card, not tested)

Card: transformers>=4.52.1. Transcribe in segments of up to ~2 minutes.

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
model_name = "ibm-granite/granite-speech-5.0-2b"
processor = AutoProcessor.from_pretrained(model_name)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_name, device_map="cuda", torch_dtype="bfloat16")

Prompt copied from the card: <|audio|>transcribe the speech with proper punctuation and capitalization. vLLM serve is also documented. The family blog names llama.cpp and mlx-audio. forkalt has not run those recipes.

Speaker tags belong on 4.1-2b-plus, not this from_pretrained id.

License

Card tag apache-2.0 on the three Apache cards named above. No LICENSE file in the 5.0-2B siblings list we opened. The turboctc-nc sibling is CC-BY-NC-SA-4.0 and is not listed as an open option here. Not legal advice.

What we did not invent

• Open ASR leaderboard / FFASR adjectives stay vendor.
• WDER / WER deltas stay self-reported.
• We do not rank these cards against Otter, Descript, or pyannote community-1.

Related compose (chat + decision + embedded TurboCTC, not this ASR family): Granite Switch 4.1. Related: Nemotron 3.5 Transcribe (NVIDIA 30B-A4B offline ASR; not Bolt/Super VL) — Nemotron 3.5 Transcribe vs Granite Speech 5.0, Hy-SimulMT (Tencent Hunyuan 30B-A3B simultaneous S2ST, different job) — Hy-SimulMT vs Granite Speech 5.0, Hibiki-Zero (Kyutai 3B streaming S2ST, fr/es/pt/de→en) — Hibiki-Zero vs Hy-SimulMT, Otter.ai alternatives, pyannote community-1, whisper.cpp, Descript alternatives, Drama 3 (TTS, not ASR). Chat weights: MiniMax M3.1 Flash, Raon-1-mini.