openclaw - 💡(How to fix) Fix TTS: Configuration recognized but speech synthesis never triggered (no TTS activity in logs) [2 comments, 2 participants]

Describe the bug

TTS is fully configured but never triggers speech synthesis on any channel (Feishu + WeChat tested). Gateway logs show zero TTS activity despite correct configuration.

Environment

• OpenClaw version: 2026.5.18
• macOS: 26.4.1 (25E253) on Apple Silicon
• Node.js: v25.9.0
• ffmpeg: 8.1.1 (present)
• Channel: Feishu (WebSocket mode) + WeChat (openclaw-weixin plugin)

TTS Configuration

{
  "messages": {
    "tts": {
      "auto": "always",
      "provider": "openai",
      "providers": {
        "openai": {
          "baseUrl": "http://192.168.88.88:8000/v1",
          "apiKey": "***",
          "model": "fish-audio-s2-pro-8bit",
          "voice": "default"
        }
      },
      "maxTextLength": 4000,
      "modelOverrides": { "enabled": true }
    }
  },
  "tools": {
    "profile": "coding",
    "allow": ["tts"]
  }
}

What works

• Direct curl to the TTS endpoint (POST /v1/audio/speech) generates valid WAV audio successfully
• STT (SenseVoiceSmall via same Mac Studio server) works — transcribes inbound voice messages correctly
• Gateway hot-reloads TTS config changes (confirmed in logs): config change detected; evaluating reload (messages.tts.*)

What doesn't work

1. [[tts:...]] inline directives — When responding to voice messages with [[tts:]], Gateway logs dispatch complete (queuedFinal=false, replies=0) — the reply is dropped entirely. When responding to text messages, replies=1 but only plain text is delivered, no audio.

2. auto: "inbound" — Receiving a voice message does not trigger automatic TTS of the text reply. Plain text is delivered.

3. auto: "always" — After switching to always, every reply is still plain text with no audio.

Gateway logs — ZERO TTS activity

Despite the [[tts:]] directives and auto: always, the Gateway log contains no TTS processing entries at all — no speech generation, no transcoding, no errors. The only TTS-related entries are config reload notifications:

[reload] config change detected; evaluating reload (messages.tts)
[reload] config change detected; evaluating reload (messages.tts.providers.openai.model, ...)

Steps to reproduce

1. Configure messages.tts with an OpenAI-compatible TTS endpoint (tested with fish-audio-s2-pro-8bit and Kokoro-82M-bf16)
2. Enable TTS tool via tools.allow: ["tts"]
3. Set auto: "always" or use [[tts:...]] in replies
4. Send a voice or text message on Feishu/WeChat
5. Observe: reply is text-only, no audio is generated or delivered

Expected behavior

With auto: "always" and a working TTS provider, all assistant replies should be delivered as voice messages (or text + voice) on channels that support native voice notes (Feishu, WeChat). With [[tts:...]] directives, the tagged text should be synthesized and delivered as audio.

Additional context

• Multiple TTS models tested: Kokoro-82M-bf16, fish-audio-s2-pro-8bit (both work via direct curl)
• Tool tts was not in the coding profile defaults — had to be manually added to tools.allow
• Setting modelOverrides.enabled: true made no difference
• Same behavior on both Feishu and WeChat channels
• WeChat conversation also showed "only text, no voice" (parallel testing, separate session)