feat(voice): backend-aware voice picker for Gemini Live and Qwen Realtime

[heavygee]

Summary

Extend the Settings voice picker (shipped in #690 for ElevenLabs) so Gemini Live and Qwen Realtime each get a per-backend voice dropdown, with selections stored independently when switching backends.

Context

• feat(voice): voice picker in settings — select from account voices including clones #686 — original voice picker issue
• feat(voice): dynamic settings voice picker with safe fallback + preview #690 — merged (ElevenLabs dynamic picker + preview)
• feat(voice): pluggable voice backend with Gemini Live & Qwen Realtime #692 — open (pluggable voice backends); this work stacks on feat(voice): pluggable voice backend with Gemini Live & Qwen Realtime #692, not bare main

Implementation plan (fork): docs/plans/voice-selection-all-backends.md
Peer handoff: docs/plans/peer-handoff-voice-selection-all-backends.md

Scope

• VoiceSessionConfig.voiceName wired through Gemini / Qwen / ElevenLabs sessions
• Per-backend localStorage keys (hapi-voice-gemini, hapi-voice-qwen, migrate from hapi-voice-id)
• Backend-aware Settings UI (static lists for Gemini/Qwen; existing /api/voice/voices for ElevenLabs)
• Hub: do not call ElevenLabs voices API when active backend is not ElevenLabs

Out of scope

• Voice preview for Gemini/Qwen (no API)
• Full Qwen 55-voice catalog until verified against DashScope docs

Merge order

1. Merge feat(voice): pluggable voice backend with Gemini Live & Qwen Realtime #692 first
2. Rebase follow-up PR onto upstream/main, retarget to main

Refs #686
Blocked by #692

[github-actions]

This is a valid follow-up feature request against the current voice settings implementation. On current main, voice selection is still wired as an ElevenLabs voiceId only, stored under one global localStorage key, so the requested per-backend Gemini/Qwen selections are not present yet.

I also did not find docs/plans/voice-selection-all-backends.md or docs/plans/peer-handoff-voice-selection-all-backends.md in the current checkout, so those plan docs appear to live outside main.

Relevant code:

• web/src/realtime/types.ts:3 - VoiceSessionConfig currently exposes voiceId?: string, not a backend-aware voiceName/per-provider field.
• web/src/lib/voice-context.tsx:43 - voice startup reads preferences from localStorage before starting a realtime session.
• web/src/lib/voice-context.tsx:46 - selected voice is read from the single hapi-voice-id key.
• web/src/routes/settings/index.tsx:342 - Settings initializes voice picker state from hapi-voice-id.
• web/src/routes/settings/index.tsx:422 - Settings writes/removes the same single hapi-voice-id key.
• web/src/routes/settings/index.tsx:432 - Settings always calls fetchVoices(api) on mount.
• web/src/api/client.ts:553 - fetchVoices() calls /api/voice/voices.
• hub/src/web/routes/voice.ts:288 - /voice/voices is implemented as an ElevenLabs voices endpoint.
• hub/src/web/routes/voice.ts:179 - /voice/token accepts voiceId and uses it for ElevenLabs token/agent resolution.
• web/src/realtime/RealtimeVoiceSession.tsx:52 - realtime session startup passes config.voiceId into the ElevenLabs token request.

Need more info:

• Please link the branch or PR that contains the two docs/plans/*voice-selection-all-backends* files, since they are not present on current main.

---

HAPI Bot