This appendix describes the transcription stage of an anonymised language education workflow. Public case material describes a cloud speech-recognition path with a local Whisper fallback; no customer audio, provider settings, or new error measurement is included.

Route by segment

File-level language detection loses information when a recording is bilingual. Choosing one language for the whole file can make the first half look plausible while damaging the other half. A single sentence can also contain a code switch, so the useful unit for routing is the reviewable segment rather than the upload as a whole.

Transcription also had to remain connected to the other pipeline outputs. A text string without timestamps or speaker labels would leave the reviewer searching the source audio again. The target was a segment record that kept its time range, anonymous speaker ID, detected language, and text together. Later checks could then point back to the same audio interval.

Keep text beside timing

The cloud path handled the normal transcription request, with a local model available when the primary path was unsuitable or unavailable. Language detection ran per segment. The result stayed alongside the boundary and diarization records rather than replacing them.

The following data model is illustrative. It describes the information the review step needs; it is not an SDK response:

{
  "start": 12.48,
  "end": 18.91,
  "speakerId": "SPEAKER_00",
  "language": "zh-en",
  "text": "..."
}

The language value may describe a mixed segment rather than a single linguistic fact. A finer-grained word-level language label could be added later, but it would also increase the amount of output a reviewer has to inspect. The segment remains the primary review unit, aligned with the script and the audio region.

When transcription paths disagree

A cloud recogniser can be convenient for the ordinary path, while a local fallback gives the workflow another option when network access, provider availability, or privacy requirements change. The trade is operational: two transcription paths can differ in punctuation, timestamps, language labels, and error patterns. Switching paths must remain visible in internal metadata, and the reviewer must be able to hear the source when text and script disagree.

Per-segment detection also has a limit. It can identify that a region contains mixed language, but it does not settle whether a phrase was transcribed correctly. Speaker assignment, script alignment, and translation completeness are separate checks. In this workflow the structured validation step flagged those mismatches; the reviewer remained responsible for deciding what to edit and export.

Evidence boundary

The current public evidence does not preserve a test set, language mix, model versions, or scoring method, so no general accuracy rate can be derived here. The supported design claim is narrower: per-segment routing and a fallback path keep mixed-language text alongside timestamps and speaker turns for human review.