Developer documentation

The Dialt voice engine

Use Dialt's API and SDKs to build voice assistants, connect your tools, run evaluations and monitor conversations. To train a model on your data, contact our team.

Get an API key Browse recipes

SDKs

Choose your SDK

Keep persistent API keys on your backend. Browser and native apps connect with scoped session credentials.

TypeScript / JavaScript SDKBrowser media, headless Node.js realtime and a generated HTTP client.npm install @dialt/sdk Python SDKServices, telephony bridges and devices that manage their own audio.uv add dialt-sdk Swift SDK · ExperimentalNative voice and text for macOS and iOS. Add dialt-ai/dialt-swift through Xcode’s Swift Package Manager.

For other languages and custom media stacks, use the WebSocket integration guide. See the Swift SDK’s validation notes for device coverage.

Using Pipecat? The Dialt-maintained Pipecat plugin (alpha) includes setup instructions, a minimal pipeline and supported features and limitations.

Recipes

Start from a working example

Dialt recipes are runnable application patterns built on the public SDKs. Each example includes its setup instructions. Your application owns its business rules and tool execution.

Agent handoffUse the Python SDK to pass a live conversation from intake to a specialist. Policy guidanceDeclare session policies and inspect guidance events with the Python SDK. Guided researchBuild an interview from a conversation plan in Python or the browser. Telnyx callsConnect your own Telnyx number through the maintained Python Call Control bridge. Twilio callsConnect inbound phone calls through the maintained Python Media Streams bridge.

API reference

HTTP APISession credentials, exact-text speech, voices, sessions and recording segments. Realtime APISession configuration, control messages, tool calls and events. Background tools guideLong-running jobs, progress, decisions, cancellation and completion.

API schemas: OpenAPI · AsyncAPI. For coding agents: documentation index.

Pre-generate audio

Exact-text speech

Render fixed text as an audio file your application stores and plays itself. Call it from your backend with a persistent key. Dialt speaks exactly what you send and keeps no copy of the text or audio. Exact-text speech is available only through POST /v1/speech; realtime session keys and WebSocket speech frames do not provide this operation.

curl https://api.dialt.com/v1/speech \
  -H "Authorization: Bearer $DIALT_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: cue-7f3a9c21" \
  -d '{"voice": "us_female", "text": "Hold for 30 seconds, then rest 90 seconds."}' \
  -o cue.wav

The response is audio/wav: mono, 24 kHz, 16-bit PCM. Send exactly one of voice, a key from GET /v1/voices, or design, the ID of one of your published Voice Designs. A voice or design private to another account is rejected with invalid_voice or invalid_design. A design contributes only its voice and delivery. There are no model, provider or tier settings.

Limits: up to 1000 characters per request, with at least one letter or digit. Longer text returns text_too_long and is never truncated; send long scripts as separate requests. Square and angle brackets return invalid_text because they are reserved for speech markup that would drop words from the audio.
Pronunciation: Dialt does not rewrite your text. The voice reads numerals, abbreviations and punctuation as it would in a live conversation, so write out anything that must be said a particular way.
Billing: the audio's length in whole seconds, rounded up, at half the Dialt Fast hourly rate (half of the saved Fast or Smart tier for a design), posted only after audio is produced. No language model runs. Refusals and synthesis failures are free. A retry with the same Idempotency-Key, voice or design, and text is billed once and returns Idempotent-Replay: true. A lost response or a 503 after settlement can still have been billed, so use the same key when retrying an ambiguous error.
Rates: an account can run a small number of requests at once. 429 and 503 include Retry-After and are safe to retry. A 402 means the account is out of credit.
Storage: store and cache the WAV yourself; every call renders again and is billed again. Responses are private, no-store.

Evaluate and improve

Learn from conversations

Session and segment metrics.delivery and GET /v1/sessions/stats's delivery pool recording-aligned speaking delivery separately for assistant and user: speaking rate in words per minute, estimated articulation rate excluding pauses, and pause fraction. Rates use total words and total full-word span seconds across measured speech, so longer spans carry their actual weight. Coverage is explicit in aligned_segments and unavailable_segments; missing measurements are null, while measured pause-free speech has zero pause totals. The HTTP reference lists the full counts, durations and pause distribution.

Delivery covers ordinary uninterrupted English, Hindi, Portuguese, Spanish and Arabic speech with current complete alignment. Chinese characters are never reported as words per minute. Gaps at least 200 ms within speech are estimated acoustic pauses; between-turn gaps are excluded. Estimated articulation rate is a proxy and does not measure intelligibility. Older, stale or unavailable alignment contributes no delivery values. Response latency boundaries are unchanged.

Inspect transcripts, recordings and tool calls in Sessions. Use conversation evals to assess behavior and compare changes to your instructions and tools.

Start with a simulation recipe, or use the Python evals client to run your own cases.