Speech (Text-to-Speech)

POST /v1/speech converts text into spoken audio. The assigned route must include a model with the audio_output capability.

For the reverse direction — turning audio into text — see Audio and STT.

Request

curl https://ai.latentkit.com/v1/speech \
  -H "Authorization: Bearer $LATENTKIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Welcome to LatentKit. Your request was routed automatically.",
    "voice": "alloy",
    "response_format": "mp3",
    "speed": 1.0
  }'

Fields

Voice names and formats are provider-specific. Because the route decides which provider executes, prefer widely supported values, or pin a specific model in the routing policy when your product depends on one exact voice.

Provider field mapping

SDK

const speech = await client.speech.create({
  input: 'Welcome to LatentKit.',
  voice: 'alloy',
  response_format: 'mp3',
});

speech = client.speech.create(
    input="Welcome to LatentKit.",
    voice="alloy",
    response_format="mp3",
)

Response

Speech responses use the unified runtime envelope (id, provider, model, policy_version_id, usage). The audio arrives as an output_audio block in content_blocks — base64 data plus its media type — not as a binary stream:

{
  "provider": "elevenlabs",
  "model": "eleven_flash_v2_5",
  "content": "",
  "content_blocks": [
    { "type": "output_audio", "base64": "SUQzBAAA...", "media_type": "audio/mpeg" }
  ],
  "usage": { "audio_output_characters": 20 },
  "cost_usd": 0.001
}

To write a playable file:

curl -s https://ai.latentkit.com/v1/speech \
  -H "Authorization: Bearer $LATENTKIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":"Welcome to LatentKit."}' \
  | jq -r '.content_blocks[0].base64' | base64 -d > speech.mp3

ElevenLabs voice models

ElevenLabs is available for your own API key (BYOK) connections. Enable the models on the connection, then add them to the route your key resolves to.

Your ElevenLabs API key must grant the Text to Speech permission, plus Models (read) so the connection health check can validate the credential. A key scoped without Text to Speech passes the health check and then fails on the first synthesis request.

Managed (Platform Access) routing for ElevenLabs is not enabled — these models route through BYOK connections only, and usage is billed by ElevenLabs to your own account. The cost_usd on a BYOK response is LatentKit's estimate of that provider charge, not a workspace credit deduction.

Managed Billing

Platform Access speech consumes workspace credits when the selected model is available for managed usage. BYOK speech is charged by your provider account. Check current availability and pricing in the console.

Audio output inside chat

Chat requests can also request spoken output by including "modalities": ["audio"] when the route model supports audio output. When present, LatentKit requires the audio_output capability for route eligibility.

Troubleshooting

• NO_HEALTHY_PROVIDER means no route model has the audio_output capability. Enable a TTS-capable model on a connection and add it to the route. A healthy connection is not enough on its own — the model must also be in the published route.
• A 400 naming response_format means the provider does not accept that format. See the field mapping above for the accepted values.
• total_tokens: 0 alongside a non-zero charge is expected for speech: the billable unit is characters.
• If a model is unavailable through Platform Access, use a supported BYOK connection or select another route model.

Related

• Audio and STT — transcription and translation
• Models & providers
• Routes & policies