Files
stemengine_old/20260316_QA_Matrix_UI_Models_Parakeet.md
T
2026-03-16 16:35:08 +09:00

3.8 KiB

QA Matrix — UI Models + Parakeet Rollout

Date: 2026-03-16 Scope: Main Separator, Project page, Auto Pipeline

Current status snapshot (2026-03-16)

Automated/implementation validations completed:

  • Backend targeted validation suite passed in runtime container (16 tests, OK).
  • Frontend quality check passed (typecheck + lint + build:local).
  • Frontend assets rebuilt and services restarted to expose latest UI changes.

Still pending for full sign-off:

  • Manual Parakeet runtime E2E checks across remaining UI surfaces.
  • Final QA sign-off and rollout approval.

Environment

  • App stack running with backend + frontend
  • Optional Parakeet sidecar profile enabled for Parakeet checks
  • Test project with at least vocals/drums/bass/other stems

A. Main Separator

A1. Per-instrument refinement UI

  • Toggle enables/disables per-instrument controls
  • Each stem selector accepts a model
  • Clear action resets mappings
  • Summary block reflects selected mappings

A2. Submission payload behavior

  • With toggle OFF, no instrument_models is sent
  • With toggle ON + selected stems, instrument_models is sent
  • Backward behavior remains unchanged when no mappings selected

B. Project Page

B1. Lyric engine selector

  • Engine selector appears in Lyrics Transcribe mode
  • WhisperX model list appears for WhisperX
  • Parakeet model list appears for Parakeet

B2. Alternate handling

  • Parakeet disables/hides WhisperX-only controls (diarize, lsync, two-pass, explicit)
  • Parakeet disables translate task
  • Parakeet unavailable warning appears when runtime is down

B3. Reprocess UX

  • Single-stem reprocess still works
  • Batch reprocess profile can select model and multiple stems
  • Batch apply queues jobs for selected stems

C. Auto Pipeline

C1. Settings

  • Per-instrument mapping editor persists values
  • Lyrics step supports engine + model selection
  • Preset save/load keeps new fields

C2. Runner

  • Separation stage sends configured instrument_models
  • Lyric stage sends configured engine + model
  • Progress text includes selected engine/model
  • Cancel/retry behavior unchanged

D. Backward compatibility

  • Existing auto steps with whisperModel still load and run
  • Existing presets without instrumentModels still load
  • Existing project lyric requests without engine default to WhisperX

E. Sign-off

  • QA pass complete
  • Known issues documented
  • Rollout approved

Verification evidence (2026-03-16)

API/runtime checks executed:

  • /api/system/health returned OK.
  • /api/separate/per-instrument-models returned model catalog (count=51).
  • /api/separate accepted legacy payload and returned job id.
  • /api/separate accepted valid instrument_models payload and returned job id.
  • /api/separate rejected invalid instrument_models JSON with HTTP 400.
  • /api/projects/{id}/lyric-transcribe with engine=parakeet, task=translate rejected with HTTP 400.
  • /api/projects/{id}/lyric-transcribe without engine returned job id (default engine backward compatibility).
  • Post-fix Parakeet E2E check (project route) succeeded with engine=parakeet, model=parakeet-1.1b-rnnt-multilingual-asr and reached job done with transcription artifacts (_transcription.json, .elrc, _lyrics.txt).
  • Post-fix Parakeet E2E check (upload route) succeeded with a real MP3 input and reached job done with transcription artifacts (_transcription.json, .elrc, _lyrics.txt).

Observed current blocker:

  • Runtime blocker cleared after sidecar alias/signature compatibility updates.
  • Remaining sign-off work is manual UI-level validation (Project page controls + Auto Pipeline UX) and rollout approval.