Skip to main content

​
Text-to-speech

For longer inputs, split on sentence/paragraph boundaries and synthesize in parallel, or use tts.stream() which chunks for you. Output format
  • tts.create / /v1/tts: WAV — 24 kHz, 16-bit PCM, mono.
  • /v1/audio/speech: wav or pcm (raw 16-bit PCM @ 24 kHz).
  • tts.stream / /v1/tts/stream: a sequence of WAV chunks (framed).
  • MP3 / Opus / µ-law are not native yet — convert with ffmpeg (roadmap).

​
Speech-to-text

Output format. JSON. Transcripts are lowercase and unpunctuated — the model’s output vocabulary contains no capitals and no punctuation, so nothing is being stripped. Punctuation and truecasing are on the roadmap. Word timestamps. Add ?timestamps=true for words and segments with start/end times in seconds — free, since the alignment comes out of the same decode pass. srt, vtt and verbose_json all work on the OpenAI-compatible endpoint. Precision is tens of milliseconds; speaker diarization is on the roadmap. See word-level timestamps. There is deliberately no language parameter — see the STT overview for why.

​
Models & voices

  • Text-to-speech: tokay-kk-v1 (Kazakh) with six voices (Aigerim default). New models/languages appear on GET /v1/models.
  • Speech-to-text: seta-kk-ru-v2 (Kazakh + Russian, code-switching). Listed by geko.stt.models().
  • Always fetch /v1/voices rather than hard-coding the catalog.

​
Rate limits

There is no hard per-second rate limit today on either product — throughput is bounded by your credit balance. Requests run concurrently; the SDK does not serialize them. (Explicit rate limiting may arrive later.)

​
Cold starts

Both GPU backends scale to zero when idle to save cost, so the first request after an idle period can take tens of seconds; subsequent calls (within a few minutes) are fast. The SDK’s default 120 s timeout is sized for this. The two services warm independently — a synthesis call does not warm the transcriber, or vice versa. If you need both hot for a live session, pre-warm both. See Latency & quality for tuning and keep-warm tips.

​
Next steps

Latency & quality

Tune the nfe dial and handle cold starts.

Streaming

Chunk long inputs to start playback sooner.

Billing & credits

Throughput is bounded by your credit balance.

Errors & retries

What happens when a request exceeds a limit.