Provenance Record

Compact Briefing (553 words, range 300-800) with Overview, per-model sections and a What We Don't Know section. Accurate, well-structured, appropriate technical depth for developers.

Read gzipped snapshots source-0.html.gz (IBM HF blog 'Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster', Published October 21, 2026) and source-1.html.gz (granite-speech-5.0-2b model card); both HTTP 200, suspicious_patterns null for both, no embedded instructions found. Then re-fetched all live pages with curl on Oct 27, 2026 (raw HTML): the blog, the 5.0-2b card, and also the 5.0-470m-turboctc and 4.1-2b-plus cards. Live blog matches snapshot on every number checked. Verified: (1) three models and wording: blog says 'a new speech-language model, granite-speech-5.0-2b, and revised versions of two models ... granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus'; summary table Status column: New / Updated / Updated; 4.1-2b-plus 'keeps the 4.1 name because the architecture, prompt format, and API are unchanged, so it's a drop-in replacement ... with improved weights' - 'revised/updated' is source wording. (2) Apache 2.0 on all three: blog states 'These models are all released under an Apache 2.0 license with native support in Transformers' (article quote verbatim) and 'The model ships under Apache 2.0' for the 2B; live cards of all three models carry license:apache-2.0 metadata (5.0-2b card also shows 'License: Apache 2.0'); title/summary claim supported. (3) Leaderboard: Chart 1 caption verbatim 'granite-speech-5.0-2b ranks #2 among top open-weight ASR models on average English WER (5.06%). ... Results are official, as of 21 October 2026, and averaged over the Open ASR Leaderboard's public English short-form test sets' - article quotes accurate and labelled as IBM-reported. (4) Turbo CTC: 19,870 RTFx on the Open ASR Leaderboard, 'up from roughly 12,600 RTFx at launch: about 1.6x faster, at the same accuracy level' (verbatim), Conformer-block compile/graph reuse, English-only 470M encoder introduced in August, seven languages (English, Japanese, Spanish, German, French, Italian, Portuguese), native punctuation/capitalization, reverb robustness - all match. (5) 4.1-2b-plus Table 1 values match exactly: AAS all speech 210.86->98.54, biased-word WER 11.79->6.48, long-form WER 13.53->10.22, WDER 4.40->4.10, cpWER 30.50->28.48; the article correctly omits the unused percentages and says lower is better. (6) 2B: successor to 4.1 2B at same parameter count (blog and card; card metadata ~2.3B params, article gives no count beyond 'same'); languages EN/FR/DE/ES/PT/JA plus translation to/from English and English to Italian and Mandarin; 390,000 hours public audio plus synthetic data for Japanese ASR, keyword biasing, translation; keyword lists up to 2,000 terms; language-ID prompting; capitalization F1 'ranges from 89.7 to 99.5 across our test sets' (verbatim); vLLM, llama.cpp, mlx-audio support (blog); model card Release Date October 20, 2026; card quote 'For best results, transcribe audio in segments of up to ~2 minutes; split longer recordings (e.g. with VAD) before decoding.' verbatim; Granite Guardian quote verbatim. Article's blog date Oct 21 correct. Both sources are IBM first-party on huggingface.co (exact hostname 'huggingface.co' on config/source_allowlist.txt line 230). Both body URLs are in article.sources and vice versa (no orphans). Article states plainly the sources contain no independent testing, and labels figures IBM's own with the Oct 27 read date.

No discrepancies found; every number, date, language list, model name and quote traces to the snapshots and the live pages. Minor observation only: 'revised' in the summary versus 'Updated' in the blog table is source wording for both.

Accurate, well-attributed first-party release briefing ready for publication; verdict APPROVE. Human-requested: no. Featured candidacy: 1 - accurate open-weight speech model release coverage for a defined developer audience, IBM first-party sourcing only