P1 — wire transcript normalization into voice, dictation and semantic routing before #122 #216

New issue

Open

opened 2026-08-28 03:23:52 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-08-28 03:23:52 +00:00

Owner

Parent: #215
Depends on: #216, #217, #218
Integrates with: #12, #25, #82/#83, #122, #132, #160

Goal

Integrate the selected post-STT transcript normalizer into Zara's actual user-input paths without creating a second voice stack or changing STT provider semantics.

Canonical placement

For voice/conversation turns:

final STT transcript
  -> transcript normalization
  -> optional #122 bounded semantic rewrite
  -> deterministic semantic routing / conversational fallback

For dictation:

final STT transcript
  -> transcript normalization
  -> ordered typing/output sink

Do not normalize partial transcripts in this issue.

Runtime behavior

normalization starts only after canonical final-transcript acceptance;
preserve turn/trace/principal/conversation correlation;
cancellation before/during normalization prevents later routing/output;
stale normalization result cannot append history, trigger intent routing or type into the dictation sink;
selected failure policy is honored exactly;
off backend preserves existing behavior;
successful normalized text is the text routed onward;
#122 receives normalized text, never bypasses this stage when both are enabled;
final runtime/client transcript events make raw-vs-normalized state explicit enough for supported UI/diagnostics without emitting two user turns.

Dictation requirements

Reuse #12 ordering/lifecycle guarantees. When parallel ASR work completes out of order, normalization and final output must still commit in capture sequence. A slow normalizer for chunk N cannot permit chunk N+1 to type first unless the dictation architecture explicitly changes to a different ordered commit mechanism with equivalent guarantees.

Daemon / client behavior

Reuse #132 and existing runtime event transport. Do not run S1-mini separately in each client when the canonical daemon owns STT/semantic processing. Client-owned display may show raw partial text and later replace/annotate the final with normalized text according to the runtime event contract.

Tests

TDD end-to-end fake pipeline cases for:

off unchanged path;
S1-style cleanup used for semantic routing;
correction changes the routed utterance to the speaker's final intended wording;
normalized dictation output preserves order;
normalizer timeout with raw fallback;
fail-turn policy;
cancel during normalization;
stale result after replacement turn;
#122 receives normalized input exactly once;
no duplicate conversation user message;
daemon event ordering raw-final -> normalization result -> route;
two principals normalize independently without shared mutable provider state.

Add one focused non-interactive integration script and include deterministic cases in full repo/Nix CI.

Acceptance

With backend = "s1-mini", real Zara voice/dictation paths use normalized final text before routing/output; with backend = "off", behavior remains compatible. Cancellation, ordering, daemon ownership and #122 placement are regression-proven.

Parent: #215 Depends on: #216, #217, #218 Integrates with: #12, #25, #82/#83, #122, #132, #160 ## Goal Integrate the selected post-STT transcript normalizer into Zara's actual user-input paths without creating a second voice stack or changing STT provider semantics. ## Canonical placement For voice/conversation turns: ```text final STT transcript -> transcript normalization -> optional #122 bounded semantic rewrite -> deterministic semantic routing / conversational fallback ``` For dictation: ```text final STT transcript -> transcript normalization -> ordered typing/output sink ``` Do not normalize partial transcripts in this issue. ## Runtime behavior - normalization starts only after canonical final-transcript acceptance; - preserve turn/trace/principal/conversation correlation; - cancellation before/during normalization prevents later routing/output; - stale normalization result cannot append history, trigger intent routing or type into the dictation sink; - selected failure policy is honored exactly; - `off` backend preserves existing behavior; - successful normalized text is the text routed onward; - #122 receives normalized text, never bypasses this stage when both are enabled; - final runtime/client transcript events make raw-vs-normalized state explicit enough for supported UI/diagnostics without emitting two user turns. ## Dictation requirements Reuse #12 ordering/lifecycle guarantees. When parallel ASR work completes out of order, normalization and final output must still commit in capture sequence. A slow normalizer for chunk N cannot permit chunk N+1 to type first unless the dictation architecture explicitly changes to a different ordered commit mechanism with equivalent guarantees. ## Daemon / client behavior Reuse #132 and existing runtime event transport. Do not run S1-mini separately in each client when the canonical daemon owns STT/semantic processing. Client-owned display may show raw partial text and later replace/annotate the final with normalized text according to the runtime event contract. ## Tests TDD end-to-end fake pipeline cases for: - `off` unchanged path; - S1-style cleanup used for semantic routing; - correction changes the routed utterance to the speaker's final intended wording; - normalized dictation output preserves order; - normalizer timeout with raw fallback; - fail-turn policy; - cancel during normalization; - stale result after replacement turn; - #122 receives normalized input exactly once; - no duplicate conversation user message; - daemon event ordering raw-final -> normalization result -> route; - two principals normalize independently without shared mutable provider state. Add one focused non-interactive integration script and include deterministic cases in full repo/Nix CI. ## Acceptance With `backend = "s1-mini"`, real Zara voice/dictation paths use normalized final text before routing/output; with `backend = "off"`, behavior remains compatible. Cancellation, ordering, daemon ownership and #122 placement are regression-proven.

nsaspy referenced this issue

2026-08-28 03:23:52 +00:00

P1 — add transcript-normalizer configuration, selection, status and diagnostics #217

nsaspy referenced this issue

2026-08-28 03:23:53 +00:00

P1 — implement local S1-mini transcript normalizer backend #218

nsaspy referenced this issue

2026-08-28 03:24:19 +00:00

[EPIC] Selectable post-STT transcript normalization backends, including local S1-mini #220

nsaspy referenced this issue from a commit

2026-09-08 02:00:55 +00:00

docs: record #216 exact-master recovery