Skip to main content
Batch and Real-time Speech to Text are available at api.myvocal.ai with your existing accessKey. Both use the workspace’s STT rates, shared Characters balance and History. This page describes the release verified on October 5, 2026.

​
Available workflows

  • Batch: multipart upload or a public mediaUrl, asynchronous submission and polling, optional bounded wait, transcript retrieval and deletion.
  • Results and exports: text, word/character alignment when requested, supported entity and text-processing options, per-channel results, and txt, json, srt, segmented_json, html, docx and pdf downloads.
  • Real-time: session creation, audio frames, partial/final results, manual or VAD commits, pause/resume within the tested session flow, explicit finish and one-use browser tickets.
  • Billing: optional processing features are included in the base STT rate. Multiple audio channels do not multiply the price of the same file timeline. Replaying a submission with the same idempotency key and unchanged input does not start another billable task.
Read capabilities for account access, rates and current option constraints. An available mode does not mean every optional feature or option combination is supported.

​
Known parameter error issue

Some invalid options.entityCategories values currently result in an asynchronous FAILED task with error.errorCode = MEDIA_UNREADABLE, instead of INPUT_INVALID. This can happen even when the audio is valid. Review the requested categories before replacing the file or submitting again. The failed transcription’s reservation is released; check billing for the actual amounts. Do not retry an invalid parameter unchanged. A corrected request is a new logical task and needs a new Idempotency-Key. General schema errors and unsupported option combinations are still rejected as input errors.

​
Completion notifications

Completion delivery is implemented. Production verification of raw-body HMAC signatures, retries after a non-success response and receiver de-duplication is still incomplete. Keep polling as the authoritative result path. See the notification contract to implement and verify your receiver. Retries are bounded; delivery is not guaranteed.

​
Real-time disconnections and recovery

Explicit finish, reading the resulting History record, and repeated finish have been verified after a service replacement, including release of unconfirmed audio’s reserved Characters. The complete disconnect → pause → resume into a new epoch flow across service replacement has not yet been verified. Do not assume uninterrupted recording or automatic recovery of missing audio. After a disconnect, read the session and its accepting state. Use the REST finish endpoint to close it when you are done; inspect PARTIAL results and usage. Never treat a closed socket as proof of successful transcription or final billing.

​
Speaker matching and storage requests

matchKnownSpeakers is not available. Speaker separation within a recording is a different feature and remains available through the documented options. processingContentStorage requests a processing-storage preference. Eligibility and the actual retention effect have not been verified. Acceptance of the request, or absence of a PROCESSING_STORAGE_NOT_APPLIED notice, is not a zero-retention guarantee.

​
Client validation coverage

Production real-time calls have been verified for PCM at 8, 16 and 48 kHz and μ-law at 8 kHz. PCM at 22.05, 24 and 44.1 kHz is implemented and locally tested, but has not yet been verified in production. The browser demo has behavior tests; real-microphone/device acceptance remains unverified. Always declare the actual sample rate and test your capture device.