standard_asr.contract.results

Per-channel transcription for multi-channel audio.

A structured, non-fatal notification from the standard layer.

Diagnostics report lossy conversions, assumed parameters, best_effort degradations, and similar non-ideal paths.

Segment-level detail, shared between batch results and streaming events.

The constant-shape result returned by transcribe and stream reduction.

The top-level text / segments / words are always the complete, channel- and speaker-agnostic transcription. For multi-channel audio they are the time-merge of all channels (never channel-0-only), so ignoring channels is always safe and lossless.

Word-level detail, shared between batch results and streaming events.

Derive a segment-level speaker label from its words (the pinned synthesis rule).

THE single synthesis rule of the standard layer, used when an engine populates Word.speaker but leaves the authoritative Segment.speaker None:

• Majority by word count -- the label carried by the most words wins.
• Tie -> the speaker of the earliest (lowest-index) word among the tied labels.
• Words with speaker=None do not vote.
• No speaker-bearing words (or words None/empty) -> None.

This function is deliberately the ONLY implementation -- both the batch post-processing (EngineBase.transcribe) and the streaming reducer (StreamReducer) call it, never a private copy. Portability demands it: the same engine and audio MUST yield the same Segment.speaker whether the app took the batch or the streaming path; two drifting copies of the rule would silently break that promise.

Project a Python value into the wire value space.

Every wire-visible slot -- Diagnostic.provided / effective, every extra mapping -- is declared JsonValue, because the Python objects and the JSON documents are meant to be the same protocol seen twice (G5.2). Declaring them Any admitted values with no JSON representation at all, which then failed during the wire projection -- after an endpoint had already committed to a response.

Two things stand between an ordinary value and that declaration, and this helper is where both are handled:

• a structured value (a pydantic submodel such as a DiarizationRequest) has a JSON form but is not itself JSON, so it is dumped;
• a typed container (list[str], dict[str, int]) IS JSON data, but a type checker does not accept it where list[JsonValue] is expected, because list is invariant. That is a static-analysis artifact, not a real mismatch, so it is absorbed here once instead of forcing a cast at every call site.

Runtime validation is unaffected: the model still validates what it is given, so a value that is genuinely not JSON is rejected loudly at construction, naming the field.

Validate a speaker label at construction (shared by every speaker field).

One rule for Segment, Word, and TranscriptionEvent: a label must be non-empty, not whitespace-only, and carry no leading/trailing whitespace. "" would be an undefined third state between None (no attribution) and a real label; edge whitespace ("A " vs "A") silently breaks within-result label consistency -- two strings = two speakers -- so both are rejected, never normalized (normalizing would hide an engine bug behind a silently rewritten value; the same stance as phrase_hints). None (no attribution) passes through.