v1.1: Orpheus TTS (tier 2) activation — weights download + TTFA verification #77

New Issue

2026-05-12T19:29:51Z

KofTwentyTwo commented

2026-05-12 19:29:51 +00:00

(Migrated from github.com)

Summary

Activate the Orpheus tier-2 TTS path. Code wiring (mlx-audio-swift v0.1.2, LlamaTTSModel.fromPretrained) is present in packages/Voice/Sources/Voice/TTS/, but the ~6 GB weights are not downloaded and the path is gated behind a MissingT2Provider-style placeholder. Until activated, Settings.setTTSTier(.tier2) is a no-op.

Evidence

Migration plan: `.planning/architecture/JARVIS-API-MIGRATION-PLAN.md:592, 731` lists Orpheus tier-2 activation as v1.1 milestone.
v1.0 spec §6 lists it with `[trigger: deferred]` annotation.
Voice audit P0-4 (#31) covers the related `tts.tier` config wire-up bug; this issue is the activation itself.

Gap classification

"design-mapped but not built" — Voice surface command `synthesizeTurn(turnId:text:tier: .tier2)` exists in v1.0 contract; the .tier2 branch never lights up in production.

Activation requirements

~6 GB Orpheus weights downloaded (or fetch-script wired).
`JARVIS_REAL_MODELS=1 swift test --package-path packages/Voice --filter OrpheusTTFATests` green.
Empirical TTFA 150–250 ms verified on Apple Silicon.

Recommended action

File against v1.1 milestone (post M-7). Depends on #31 (`VoiceTTSAdapter.tierResolver` wired) closing first.

Cross-references

Related: #31 (tts.tier config dead), #50 (B-05 TTS silent)
Migration plan §10 v1.1

## Summary Activate the Orpheus tier-2 TTS path. Code wiring (`mlx-audio-swift v0.1.2`, `LlamaTTSModel.fromPretrained`) is present in `packages/Voice/Sources/Voice/TTS/`, but the ~6 GB weights are not downloaded and the path is gated behind a `MissingT2Provider`-style placeholder. Until activated, `Settings.setTTSTier(.tier2)` is a no-op. ## Evidence - Migration plan: \`.planning/architecture/JARVIS-API-MIGRATION-PLAN.md:592, 731\` lists Orpheus tier-2 activation as v1.1 milestone. - v1.0 spec §6 lists it with \`[trigger: deferred]\` annotation. - Voice audit P0-4 (#31) covers the related \`tts.tier\` config wire-up bug; this issue is the activation itself. ## Gap classification \"design-mapped but not built\" — Voice surface command \`synthesizeTurn(turnId:text:tier: .tier2)\` exists in v1.0 contract; the .tier2 branch never lights up in production. ## Activation requirements 1. ~6 GB Orpheus weights downloaded (or fetch-script wired). 2. \`JARVIS_REAL_MODELS=1 swift test --package-path packages/Voice --filter OrpheusTTFATests\` green. 3. Empirical TTFA 150–250 ms verified on Apple Silicon. ## Recommended action File against v1.1 milestone (post M-7). Depends on #31 (\`VoiceTTSAdapter.tierResolver\` wired) closing first. ## Cross-references - Related: #31 (tts.tier config dead), #50 (B-05 TTS silent) - Migration plan §10 v1.1

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: KofTwentyTwo/Jarvis#77