Description

When using Talk Mode on the macOS companion app with a non-ElevenLabs TTS provider (e.g. system), two problems occur:

1. Double playback: Every assistant reply is spoken twice back-to-back
2. Premature cutoff for CJK languages: The watchdog timer uses a flat 0.08s/char estimate that is too short for Korean, Chinese, and Japanese text, causing speech to be killed mid-sentence

Environment

• OpenClaw: 2026.3.13
• macOS: Darwin 25.2.0 (Mac mini)
• Talk provider: system (also reproduced with openai)

Steps to Reproduce

Double playback

1. Set talk.provider to system (or any non-ElevenLabs provider)
2. Enable Talk Mode in the Mac app
3. Speak a message and wait for the assistant reply
4. Result: Audio plays twice back-to-back

CJK cutoff

1. Set Talk Mode language to Korean (ko-KR)
2. Ask a question that produces a 50+ character response
3. Result: Speech stops mid-sentence after ~4 seconds (50 × 0.08 = 4s watchdog)

Root Cause

Double playback

In TalkModeRuntime.playAssistant(), the error handling always falls back to playSystemVoice() regardless of which TTS provider failed:

do {
    if apiKey != nil && voiceId != nil {
        try await self.playElevenLabs(...)
    } else {
        try await self.playSystemVoice(...)  // 1st play
    }
} catch {
    // Always falls back to system voice, even when system voice itself failed
    try await self.playSystemVoice(...)  // 2nd play (duplicate!)
}

CJK cutoff

The watchdog uses 0.08s/char for all languages, but CJK characters represent full syllables:

Source: Pellegrino et al., Science Advances (2019)

Fix

1. Only fall back to system voice when ElevenLabs fails — not when system voice itself fails
2. Use language-specific per-character estimates with 3x safety margin for the watchdog

Related Issues

• Talk Mode: ElevenLabs audio playback fails immediately, falls back to system voice (macOS) #15460 — Talk Mode: ElevenLabs audio playback fails immediately, falls back to system voice (macOS) — documents the ElevenLabs → system voice fallback path where this double-play bug occurs
• TTS tool result causes duplicate audio delivery — model re-echoes MEDIA path #17991 — TTS tool result causes duplicate audio delivery — model re-echoes MEDIA path — similar duplicate audio symptom but different root cause (tool result re-echo vs fallback retry)
• Control UI webchat: duplicate assistant messages rendered on every reply #5964 — Control UI webchat: duplicate assistant messages rendered on every reply — client-side duplicate rendering that compounded with the TTS duplicate
• [Bug]: Telegram duplicate messages - text and audio sent twice #30316 — Telegram duplicate messages: text and audio sent twice — related duplicate audio delivery in channel context
• Talk Mode: Support on-device TTS (iOS AVSpeechSynthesizer) as alternative to ElevenLabs #42630 — Talk Mode: Support on-device TTS (iOS AVSpeechSynthesizer) as alternative to ElevenLabs — on-device TTS support request; the system voice path fixed here is the macOS equivalent
• Feature Request: Native macOS TTS (NSSpeechSynthesizer) support for Talk Mode #9160 — Feature Request: Native macOS TTS (NSSpeechSynthesizer) support for Talk Mode — original request for native macOS TTS support