Voice APIs
Voice APIs should measure time to first playable audio separately from time to completion. Connection setup, queueing, synthesis, transfer, decoding, and playback can each dominate a different workload.
Review current samples, pricing, limits, and documentation before production use.
What this page helps you evaluate
Build a speech integration with explicit timing, reliability, audio-format, and cost boundaries.
Integration path
Measure the whole speech request, not one vague latency number
Send a validated request over the documented endpoint.
Track acknowledgement, first audio, completion, and failure.
Decode the documented format and apply retry or fallback policy.
Production readiness
- Keep API keys out of source code and browser bundles.
- Choose compatible voice and model identifiers.
- Measure warm and cold paths separately.
- Log request IDs and sanitized timing data without user text.
Developer brief
Design Voice APIs around observable boundaries
Voice APIs should measure time to first playable audio separately from time to completion. Connection setup, queueing, synthesis, transfer, decoding, and playback can each dominate a different workload.
Use the documented REST or WebSocket contract, protect API keys, validate responses, and record percentiles and errors—not a single best-case request.
Request contract
Validate model, voice, text, language, and output settings for Voice APIs. Run the initial Voice APIs trial with a fixed script and settings so later voice or model changes remain comparable. Review Voice APIs on the actual playback device and connection profile instead of relying only on a studio headset.
Timing model
Timestamp connection, acknowledgement, first playable audio, and completion. Use a compact Voice APIs scorecard with intelligibility, pronunciation, pacing, fit, and correction effort rated independently. Attach corrections to the exact Voice APIs script segment so the team can distinguish content edits from delivery edits.
Failure path
Use bounded retries, idempotent behavior where supported, and clear fallbacks. Monitor corrections and rejected output for Voice APIs; a rising review burden can matter before a technical failure appears. When Voice APIs fails its acceptance check, retain the request metadata and sanitized timing—not sensitive source text—in the incident note.
What the product currently documents
Fixed public voice samples are available for review before purchase.
Samples are fixed previews, not a free custom-generation endpoint.
Review sourceCurrent plans, balances, rates, limits, and commercial terms are published on the pricing page.
Pricing can change; use the linked page as the current source.
Review sourceThe current public Pay As You Go plan lists 2 concurrent requests.
Plan limits can change; verify the linked pricing page before deployment.
Review sourceThe documented v3 API supports asynchronous text-to-speech generation and status tracking.
Reviewed 2026-07-24.
Review sourceThe documented WebSocket endpoint streams live raw PCM float32 mono audio at 24 kHz for supported models.
This is an audio-format contract, not a numeric latency claim.
Review sourceQuestions specific to voice apis
What should a low-latency benchmark report?
Report time to first playable audio and full completion separately, with percentiles, errors, workload, region, and connection state.
Can the WebSocket chunks be treated as MP3?
No. Follow the current streaming documentation for the live raw PCM format and use the completed file URL when appropriate.
Where should an API key be stored?
Keep it in a server-side secret store or environment configuration, never in source code or client-side JavaScript.
Test Voice APIs with your own acceptance criteria.
Review current samples, pricing, limits, and documentation before production use.