Local Voice Reply

Use this skill to turn text into a cloned/custom-voice audio reply and deliver it reliably to Feishu or Discord.

Structured skill definition

• - Purpose: local low-latency voice replies in Opus/Ogg.
• Channels: Feishu + Discord.
• Default voice: juno (reference file: voice/juno_ref.wav).
• Custom voice modes:

• - Output: .opus (Ogg container) under .openclaw/media/outbound/voice-server-v3/ (or TARVIS_VOICE_OUTPUT_DIR).
• Control scripts:

Server implementation is kept with the skill (not workspace root):

• - server/voice_server_v3.py (FastAPI routes)
• INLINECODE11 (generation and cache engine)

Voice assets are also colocated with the skill:

• - INLINECODE12

Runtime requirements

• - ffmpeg must be installed and available on PATH (required for Opus encoding).
• Python packages required by the server:

• - On first startup, ChatterboxTTS.from_pretrained() may download model assets, so initial run can require network access and additional disk.
• Optional env vars:

Persistence behavior

• - Uploaded voice samples from POST /voice/register are persisted under server/voices/.
• Cache and registry data are persisted under server/voice_cache/.
• Generated Opus outputs are written under .openclaw/media/outbound/voice-server-v3/ by default (or TARVIS_VOICE_OUTPUT_DIR when set).
• INLINECODE34 only deletes staged .opus files inside the configured output directory and their .json sidecar files.

Use this workflow

1. 1. Ensure local v3.3 TTS server is running from this skill folder:

1. 2. Call /speak with text (and optional speed, exaggeration, cfg).

1. 3. Receive Opus directly from server (audio/ogg) in Juno voice.
2. Save final media into allowed path:

1. 5. Send with message tool:

Voice customization guide

A) Replace default Juno reference

1. 1. Replace voice/juno_ref.wav with your target reference voice sample.
2. Keep sample clean (single speaker, low noise, clear pronunciation).
3. Restart server and test with voice_name=juno.

B) Register additional named voices

1. 1. Call POST /voice/register with a reference sample and target voice_name.
2. Confirm registration under server/voices/.
3. Generate with that voice_name in /speak or /speak_stream.

Defaults

• - voice_name: INLINECODE62
• INLINECODE63: INLINECODE64
• Output format: Opus in Ogg container from server /speak (no post-conversion)
• Discord compatibility: Ogg/Opus is supported and can be sent as voice/audio with INLINECODE66

Speed Improvements In This Version

• - Caches model capability lookups once at startup.
• Uses torch.inference_mode() during synthesis to reduce overhead.
• Reuses phrase cache for both /speak and /speak_stream.
• Improves chunking behavior for long CJK text to avoid oversized chunks.
• Keeps latency metrics for benchmarking and tuning.

Common failure and fix

• - Error: INLINECODE70
• Fix: copy the file into .openclaw/media/outbound before sending.

Script

Use scripts/send_voice_reply.ps1 to generate Opus directly with defaults (voice_name=juno, speed=1.2).
It auto-selects /speak_stream for longer text (or when -Stream is passed) for better throughput.

For stable CUDA generation command patterns under stricter exec approval policies, use:

• - INLINECODE77