ADR-130: Provider-agnostic speaker-name recovery from episode metadata

• Status: Accepted
• Date: 2026-07-26
• Authors: Marko Dragoljevic
• Related RFCs: —
• Related PRDs: — (#876 speaker quality, #1190 corpus reprocess, #1178/#1179 ASR lock)
• Related ADRs: ADR-134 (shared core vs per-provider strategy), ADR-123 (coverage gate)

Context & Problem Statement

Every ASR mistranscribes proper nouns some of the time — they are rare, out-of-vocabulary tokens that carry no sentence context. This is not specific to any one model:

• OpenAI Whisper (the v2.2 transcripts) rendered Kevin Roose as "Kevin Russo" and Casey Newton as "Casey Noon".
• DGX-local turbo (the v2.3 transcripts) rendered them "Kevin Roos" / "Casey Noon", and guests like David Duvenaud as "David Duvino", Sebastian Malaby as "Sebastian Maliby".

Because our speaker naming reads names out of the transcript (self-introductions, host introductions), a mistranscribed name is published as the speaker label. Measured on the v2.3 turbo corpus, only 75% of published names were exactly correct; ~20% more were mangled-but-fixable and ~5% were unrecoverable (mostly garbage, not real names).

The correct spelling is almost always already in the episode metadata. Across the 90-episode corpus: 100% of episodes have a title AND a description, and 93% name at least one guest in them, spelled correctly ("My guest today is Brian Chesky…", title "Brian Chesky – AI Founder Mode"). The feed blurb likewise states the hosts ("journalists Kevin Roose and Casey Newton").

We already exploit half of this: _canonicalize_to_known_host snaps a mangled name to a configured host (exact/near first name + soundex/edit-distance surname). It lives in the shared core and runs for every provider — which is exactly why it fixed OpenAI Whisper's own "Kevin Russo" → "Kevin Roose" on the Deepgram/community-1 v2.2 corpus. The gap is that we do not apply the same matching to guests against the metadata-stated guest names.

Two defects, both provider-agnostic:

1. No guest-metadata canonicalization. Host manglings recover; guest manglings do not, even when the correct spelling sits in metadata_named / detected_guests.
2. known_hosts is empty on the full/transcription path for org-authored feeds. The transcription consumers (transcription.py) DO set job.feed_hosts = cached_hosts before the roster runs — so the anchor plumbing is intact. The defect is upstream: cached_hosts itself is empty. The host detector on this profile is gemini.detect_hosts, which short-circuits on the RSS author tag — when feed_authors is present it returns set(feed_authors) verbatim and never reads the description. Every one of the corpus's feeds is org-authored (The New York Times, NPR, The Wall Street Journal…, Financial Times…), so it returns the org, which is_network_or_org_author then strips → cached_hosts empty. Meanwhile the descriptions DO name the hosts ("Each week, journalists Kevin Roose and Casey Newton explore…"). The relabel_only / rediarize_only paths avoid this because they read hosts from sibling metadata via the deterministic detect_hosts_from_feed, which reads the description statement first. (Evidence: cached_hosts / tried.known_hosts empty on 90/90 v2.3 episodes; all 9 feeds org-authored; the deterministic parser recovers real hosts for 6/9 of them.) Nothing to do with which ASR ran.

Correction. An earlier revision of this ADR (and the reverted commit 263b1430) claimed the cause was an un-wired feed_hosts argument on the full path. A Fable-5 review showed that fix was dead code — transcription.py already overwrites job.feed_hosts from cached_hosts before it is read. The real cause is the empty cached_hosts above.

Decision

Speaker-name recovery from metadata is a provider-agnostic, shared-core capability, not a per-ASR or per-diarizer patch. Three changes, all in (or feeding) the shared naming core — none scoped to turbo:

1. Metadata name extraction is a shared input, unconditional of provider. known_hosts (feed blurb) and the stated guests (metadata_named / detected_guests, from title + description) are read the same way for every ASR and every diarizer. They already are; this ADR records it as a contract, not a scenario-specific step.

2. Extend canonicalization symmetrically to guests. Add a guest counterpart to the existing host snap: a mangled published name that does not match a known_host is snapped to a stated guest (metadata_named ∪ detected_guests) by the same fuzzy rule (exact-or-near first name + surname within soundex OR a small edit distance). It runs in the shared core after all naming paths converge (near the final plausibility gate, ADR-134), for every provider — so the v2.2 Deepgram/community-1 corpus benefits from it too, not only turbo. Guards (added after the Fable-5 review, see Invariants): an exact match to a stated ref is never re-snapped; a non-host voice is never snapped onto a known-host's spelling (preserves the host-candidate gate); corroborated refs are matched before the un-corroborated metadata_named subjects.

3. Fix the empty-cached_hosts cause: deterministic host fallback. When the LLM detector yields no host after org-author stripping, detect_feed_hosts_and_patterns falls back to the deterministic detect_hosts_from_feed, which reads the host statement out of the description first — the exact function the relabel_only / rediarize_only paths already use via sibling metadata. This populates cached_hosts, which the transcription path already threads onto the roster. One place, provider-agnostic; anchors the roster for every full reprocess AND first-pass ingest, independent of ASR or diarizer. (The earlier tuple-threading approach was reverted as dead code — see the correction note in Context.)

Invariants (unchanged — this ADR must not weaken them)

• Never author a name. A name is only ever snapped to one stated in the episode's own metadata (host blurb, title, description). We do not invent, and we do not snap to a global name list. A mangling too far from any stated name (edit distance beyond threshold, first name unrecognisable) stays unrecovered — a raw SPEAKER_NN or the mangled token, per "a wrong label is worse than an unnamed voice" (#876).
• One name, one voice. A stated name already bound to one voice is not reused for another.
• Ambiguity abstains. If a mangled name is near two different stated people, keep it unresolved.
• Exact matches are never re-snapped. A published name that already equals a stated ref is left alone (else two stated people sharing a first name could move an exact match onto a near-ref).
• The host-candidate gate is preserved. A non-host voice is never snapped onto a known-host's spelling — a guest self-introducing "Kevin Ross" is not painted host "Kevin Roose" merely because the host voice was left unnamed. Host-identity canonicalization stays gated to host-candidate voices; a genuinely mangled co-host already carries role == "host" and still snaps.

Accepted bounded risk

The guest snap matches a mangled name against metadata_named, which includes people the episode is about but not necessarily in the room (e.g. a lawsuit defendant named in the show notes). A voice that self-introduces as a genuinely different person whose name is fuzzy-close to such a subject (e.g. "Sam Alton" → "Sam Altman") could be mislabelled. We accept this bounded tail because metadata_named is required for real recoveries (e.g. "David Duvenaud" is present only there, not in the corroborated detected_guests). It is mitigated by matching corroborated refs first and by the respell-only fuzzy bound; measured incidence on the real corpus was zero (see Evidence).

Non-Goals

• Not an ASR change. We do not switch off turbo, re-transcribe, or tune the model. This is downstream / relabel-cheap.
• Not turbo-specific. Nothing here keys on dgx_whisper_model or the transcription provider.
• Not name invention. Guests with no metadata reference anywhere (a name that appears only, mangled, in the transcript) are not recoverable and are out of scope — that residual is the ASR's true, separate cost.
• Does not close turbo's coverage gap. Turbo also surfaces fewer names than OpenAI (weaker self-intro/greeting transcription → some voices never get a candidate name at all: 148 vs 183 published). That is a distinct ASR-quality issue, tracked separately, not addressed here.

Evidence

Measured on the real v2.3 turbo corpus (90 episodes), re-resolving from the frozen turbo diarization. Two separate levers, measured separately (an earlier revision reported a single "70%→88%" number against a name-in-refs oracle that was blind to wrong-voice attribution — that oracle is not used here).

Lever 1 — host anchor (deterministic fallback). The pipeline's cached_hosts was empty on 90/90 episodes (every feed org-authored → the LLM detector's org short-circuit stripped to nothing). The deterministic detect_hosts_from_feed recovers real hosts from the description for 6 of the 9 feeds (Hard Fork → Kevin Roose + Casey Newton; Unhedged → Katie Martin + Robert Armstrong; Invest Like the Best → Patrick O'Shaughnessy; No Priors → Elad Gil + Sarah Guo; The Daily → Michael Barbaro et al.; The Journal → Jessica Mendoza + Ryan Knutson). The other 3 (NVIDIA AI Podcast, Planet Money, Latent Space) name no host in their blurb and correctly stay empty — no invention.

Lever 2 — guest snap, measured attribution-aware. Isolating the snap (host anchor ON in both arms, only _recover_stated_names toggled) across all 90 episodes, and classifying every name it changes as a respell (same person, surname edit/soundex/stem-close — a mistranscription corrected) or a swap (a different person — the accepted-risk failure mode):

snap effect (90 episodes)	count
names changed by the snap	10
— respell (same person, correctly re-spelled)	10
— cross-person swap (mislabel)	0

Examples: Sebastian Maliby → Sebastian Mallaby, Nick Allardyce → Nick Allardice, RJ Skirinj → RJ Scaringe, David Duvino → David Duvenaud. The snap is conservative (10 changes corpus-wide) and made zero cross-person mislabels — the accepted metadata_named tail risk did not materialise.

Host recovery is directly proven: re-resolving a Hard Fork turbo episode with known_hosts present snapped "Kevin Roos"→"Kevin Roose", "Casey Noon"→"Casey Newton", and the guest "David Duvino"→"David Duvenaud".

Consequences

• The whole corpus improves, not just turbo. Applied provider-agnostically, the guest snap also repairs any names OpenAI Whisper mangled on the v2.2 Deepgram/community-1 corpus, and the host fallback fixes any org-authored feed on any provider.
• Cheap and reversible. Both levers are shared-core naming logic + a feed-detection fallback; the guest snap runs under relabel_only (no audio, no ASR), so the existing corpus is repaired without re-transcription. The host fallback takes effect on the next full reprocess.
• The deepgram frozen base stays stable. Per ADR-134, the guest snap is a shared-core tightening; the full 154-test diarization suite (incl. the golden-fixture arbiter) is green, and the snap is reference-bounded (it may only re-spell to a stated name, never invent or demote).
• Not yet validated end-to-end on a live full run. Both levers are proven by unit tests + deterministic re-resolution of the frozen corpus; a live full --reprocess (which re-transcribes and calls the LLM) has not been run. That is the remaining validation gap.