Configure
Voices
Pick the voice your agent speaks with: a curated set per provider, native voices for each language, samples you can hear before you have a key, and a way to change it during a call.
Where you choose a voice
- Settings page: run
sleeptalk settings --open, then open Voice, or open a service under Speech service. Each voice has a play button. - An agent's key request: the link opens the settings page on just that request, on the host or your phone. After the key, the voice step shows the same picks for the service you chose. An agent can pin a voice with
--tts-voice-id; the page then checks exactly that one. - During a call: open the gear menu, then Voices. The list holds the voices of the service speaking in that call, in the call's language.
- Terminal setup:
sleeptalk settings --terminalwalks through the same choice.
A voice change during a call is tried on one sample line first. If the provider refuses it, the call keeps its current voice and says why. A voice that works is saved to the profile.
These in-call voice controls describe Cascade. OpenAI voice models have separate controls: GPT Live takes voice and language changes on the next call and disables standalone voice previews. See voice modes.
Hear a voice before you have a key
Sleeptalk ships a short recorded sample for every voice listed below, so the play buttons on the settings page work with no key and no network request. The page plays samples in its own language only: on a page in Spanish you hear the Spanish voices speaking Spanish, and no English preview appears.
Once a service's key is saved, the play button on that service can also speak a line in any of its voices through your key, including voices without a recorded sample.
Curated voices in English
Curated picks per provider, with the default first. The default is what a new profile starts with.
| Service | Voices |
|---|---|
| Cartesia | Skylar (default), Daniel, Gemma, Parker, Archie |
| Deepgram | Hannah (default), Kit, Alexis, Cliff, Sienna |
| ElevenLabs | Brittney (default), Mark, Jessica, Eric, Sarah |
| OpenRouter | Eve (default), Ara, Rex, Hannah, Kit |
| Venice | Eve (default), Ara, Rex, Leo, Sal |
| Groq | Troy (default), Hannah |
| Together | Skylar (default), Daniel, Tara, Sky |
| Local | Heart (default, Kokoro), 28 English Kokoro voices; Lessac (Piper); Bria and Loona (Pocket TTS) |
Local also supports custom Pocket reference voices. See Local speech.
Native voices in other languages
When you choose another language, a voice that was a starting pick moves to that language's first native voice. A voice you pasted in yourself stays. The first two in each cell were measured with a round trip (a developer sentence spoken, then transcribed back); the rest follow the provider's own ranking.
| Language | Cartesia | Deepgram | ElevenLabs |
|---|---|---|---|
| Spanish (Spain) | Marta, Darío, Iria, Rafael, Celia | Alvaro, Diana, Agustina, Néstor, Carina | Carolina, David Martin, Raquel, Carmelo, Carolina Ruiz |
| Spanish (Mexico) | Fernanda, Mateo, Ximena, Ramon, Laura | Luciano, Olivia, Selena, Aquila, Javier | Marlene, Enrique Mondragón, Ana María, Enrique M. Nieto, Regina |
| French | Inès, Mathis, Zoé, Henri, Jade | Hector, Agathe | Lucie, Nicolas, Hélène, Nicolas Petit, Victoria |
| German | Marlene, Moritz, Klara, Clemens, Jennifer | Viktoria, Julius, Aurelia, Elara | Susi, Otto, Lea, Ben, Carola |
| Italian | Elena, Alessio, Liv, Fabio, Sofia | Cesare, Melia, Cinzia, Dionisio, Demetra | Violetta, MarcoTrox, Giusy, Marco, Linda Fiore |
| Portuguese (Brazil) | Helena, Felipe, Eloá, Rafael, Alice | none | Keren, Matheus, Raquel, Will, Roberta |
| Dutch | Fleur, Thijs, Noa, Stijn, Anneliese | Sander, Rhea, Beatrix, Lars, Cornelia | Roos, Ido, Emma, Marcèles, Marianne |
| Japanese | Naoki, Yuzuki, Aiko, Yuto, Sakura | Izanami, Fujin, Ama, Ebisu, Uzume | Sakura, Otani, Morioki, Kozy, Kyoko |
| Korean | Minji, Jaewon, Yeji, Taehyun, Haeun | none | Anna Kim, Hyuk, Jiyoung, Chris, Seulki |
| Hindi | Neha, Rohan, Siya, Kabir, Anika | none | Zara, Raju, Sia, Viraj, Simran |
| Chinese (Mandarin) | Jian, Hua, Lan, Liu, Shuwen | none | Coco Li, James Gao, Evan Zhao, Stella Gu |
Deepgram has no Portuguese, Korean, Hindi, or Chinese voices, so it keeps its English voice for those languages. ElevenLabs voices come from its public library and work from an ordinary key without adding them to your account.
The Grok voices on OpenRouter and Venice (Eve, Ara, Rex, Sal, Leo) speak every language, and each has a sample recorded in each one. For Chinese on OpenRouter, the recommended speaking model is Qwen TTS Flash with its own voices; see Language support.
Groq, Together, and local speech have no voices recorded in other languages. On a page in another language their voice list says so; local voices speak English only.
Use any other voice
Every service's voice list ends with a field for a voice ID. Paste an ID from the provider's library and press Enter; Browse voices opens that library. Voice IDs are one word of plain characters.
[tts]
voice_id = "db6b0ed5-d5d3-463d-ae85-518a07d3c2b4"For Deepgram the voice names its model: an aura-2-… voice uses Aura-2, a flux-… voice uses Flux.