Repeated replies now play from memory, free of charge
October 6, 2026 · platform
If your agent says the same thing more than once, a greeting or a standard "one moment please" for instance, the agent runner now keeps the synthesised audio and plays the repeat straight from memory. The delay before the caller hears it drops to close to zero, and the speech engine does less work on busy runners.
Replays are free
A reply played from memory is not billed. VoiceThere voices charge TTS credits only for text that is actually synthesised, so a cached greeting costs credits the first time and nothing after that. The same goes for cloud TTS vendors such as OpenAI or ElevenLabs: a replay never reaches the vendor, so it does not show up on your vendor bill either.
How much is kept
| Limit | Size | Speech |
|---|---|---|
| All cached replies | 16 MiB | about 87 seconds in total |
| One reply | 2 MiB | up to about 11 seconds |
Audio is stored the way it is played to the caller (48 kHz stereo, 16-bit), which works out to roughly 190 KB per second of speech. Replies longer than about 11 seconds are never cached and are synthesised every time.
There is no fixed expiry. When the memory is full, the reply that was played least recently makes room for the new one, and a deploy starts with an empty cache. Only replies that were fully synthesised are kept: a reply cut short by the caller is not. Entries belong to one project and one voice, so a change of voice or model gets fresh audio.
Turning it off
The cache is on by default. If you want every reply synthesised fresh, say for a load test of speech capacity or because you would rather not keep spoken text in memory, switch it off and deploy again.
- Docs: Phrase cache, Speech pricing
- Dashboard: Advanced voice pipeline, then TTS playback, then Phrase cache
- CLI:
voicethere projects voice-advanced set tts.phraseCache false