hermes - 💡(How to fix) Fix feat(tts): serial queue to prevent overlapping audio playback [1 comments, 2 participants]

Problem

When voice.auto_tts is enabled (CLI or gateway), TTS audio is generated for each agent response. If the agent generates a new response before the previous TTS audio has finished playing, two audio clips overlap — effectively two voices talking at once.

This happens because the current TTS pipeline has no notion of a playback queue: each text_to_speech call produces a new file independently, with no awareness of whether a previous clip is still being consumed by the client.

Expected behavior

TTS output should be serialized — a new clip should not start playing (or being sent to the client) until the previous one has completed. In practice this means:

1. The gateway/CLI tracks TTS playback state per session/chat
2. If a new agent response arrives while audio is still playing, queue it
3. Play the queued clip after the current one finishes

Affected platforms

• CLI (local audio playback)
• Gateway platforms that deliver voice bubbles (Telegram, Discord, Matrix, etc.)

Proposed approaches

• Gateway: Add a per-chat TTS queue in gateway/run.py / base.py — buffer audio files and send them sequentially with a small gap.
• CLI: Add a playback tracker in the CLI audio handler that defers sending new audio until the current one finishes.
• Config option: voice.tts_queue: true (default on) to enable/disable.

Workaround

Currently users can only mitigate this by using slower TTS providers or shorter agent responses. No configuration exists to prevent the overlap.