ASR Pipeline v2

Goal

The pipeline recovers useful meeting facts and traces rather than pretending that speech can always be reconstructed word for word. The original transcript remains the audit source. A draft may omit low-value repetition, but must not invent a subject, number, decision, or relationship.

Stages

S0 parses speaker turns and collapses decoder loops.
S1 generates candidates from the untouched transcript. Phonetics, observed ASR errors, aliases, entity relationships, and recency are independent signals.
S2 only removes fillers and repairs punctuation. It receives no Hindsight context and makes no entity decisions.
S3 applies settled mappings deterministically. Flash Lite sees only unresolved candidates and Chinese-number spans.
S4 extracts source-labelled evidence cards per chunk.
S5 Sonnet groups those cards by topic and writes a rough draft.

Personal fact store

state/lexicon.db is intentionally reusable outside ASR:

entity, alias, correction: identity and observed surface forms.
relation: person/company/project relations with source, confidence, and dates.
interaction: contact recency and frequency.
observation: what a transcript emitted and how it was resolved.
phonetic_confusion: confirmed personal accent/ASR confusion pairs.
review_question: transport-independent human review queue.
source_document: incremental indexing state for historical minutes.

Obsidian remains a curated source. Hindsight supplies broad recall. SQLite holds small deterministic facts that applications can safely query.

Human review loop

The pipeline never blocks waiting for an answer. Ambiguous high-impact entity candidates are queued and the draft keeps an uncertainty marker.

bin/asr-refine review pending
bin/asr-refine review answer 12 万亚坤

A confirmed candidate is learned as an alias, correction, and observation. The candidate-pool cache includes the database timestamp, so the next recording uses it immediately.

Pi, Telegram, and the web UI are thin transports over this queue:

Read pending rows and display the question, source sentence, and choices.
Submit question_id + answer.
Call the same review.answer() function or CLI command.
Optionally request a replay of S3-S5 for the affected recording.

The production web transport is a persistent mobile control plane:

bin/asr-refine review serve       # normally started by systemd
bin/asr-refine review state       # print the active one-time URL

Every generated draft is a Web delivery, even when entity candidate detection finds no ambiguity. Review order is: inferred call background (with a corrected background field that requests a redraft), unnamed voiceprints, the draft's ## 待你确认 list, then one overall review item.
It listens on 192.168.50.145:8765. / accepts streamed audio uploads, shows recent job state, and atomically hands completed files to the existing recording watcher. /hook/intake exposes the same raw-upload contract for a future reverse proxy or external uploader.
Telegram receives a 256-bit capability URL. The token exists only in Telegram and the mode-0600 runtime state file; it is redacted from service/request logs.
The page exposes only one queued question at a time: candidate buttons, an explicit "无法确认", a short custom answer, and an expandable full-draft view.
Rough drafts live under state/drafts/, outside the Vault and Git. After all questions are answered, an auditable feedback block is saved privately. asr-review-revise@.service applies those answers with Sonnet, validates the formal schema/Summary/provenance, then writes and pushes exactly one formal note under minutes/. The private reviewed source moves to state/published/.
A confirmed unnamed speaker_N is written to voiceprint_speakers, marked for retry by the daily enrollment job, and immediately submitted to the persistent :10099/voiceprints/enroll API.
SQLite records one entity_mention per entity per formal minute. When the current minute would bring an uncarded person/company to two distinct minutes, Web offers a conservative card scaffold. No card or judgment field is created without an explicit human choice.
Upload metadata (title and optional user-supplied background) is kept separate from the acoustic transcript. The title survives into the task slug; the background is explicitly labelled as user-supplied context for drafting.
asr-review-web.service is enabled at boot and restarts on failure. A successful private draft ensures it is running and sends its stable review URL through Telegram. dry-run never queues questions.

The service unit lives at deploy/asr-review-web.service. Telegram may later gain inline choice buttons, but neither transport owns facts or writes SQLite directly.

Confidence policy

Confirmed: voiceprint, explicit human mapping, or multiple independent strong signals.
Probable: context and relation graph agree, but identity is inferred.
Pending: weak or conflicting evidence, including numbers whose type/unit is not clear.

Only confirmed facts become deterministic aliases automatically.

Graph and recency

bin/asr-refine graph sync
bin/asr-refine graph stats

The graph incrementally scans finalized minutes, imports person/company/project frontmatter and known wikilinks, records meeting interactions, creates low-confidence co-occurrence edges, and promotes explicit linked relations such as 投资/负责/合作 at a higher confidence. Deleted or changed notes remove only facts owned by that source.

Candidate context weight combines relation confidence, repeated support, last contact, and a 45-day half-life. It influences ranking but cannot activate the personal phonetic-confusion channel outside entities already in the current graph context.

Numbers

The deterministic number pass classifies percentage, date, money, quantity, and project/code speech. Every decision records raw value, normalized value, type, confidence, source span, and nearby context. Ambiguous forms such as 两万五九份 remain unchanged and are passed to the uncertainty stage instead of being guessed.

6.2 KiB Raw Blame History