Live captions fail in subtle ways: a proper noun is rewritten seconds after the viewer reads it, a short reply loses its speaker, or the WebSocket reconnects and repeats a word. Soniox’s real-time STT documentation explains why correct token handling matters at least as much as the displayed language count.
Version and transport as of October 2026
Soniox lists `stt-rt-v5` as an active realtime model and `stt-async-v5` as its separately active recorded-audio model. The company’s version history says realtime v5 replaced v4 on June 16, 2026; v4 may act as an alias rather than an immutable model. Real-time audio is streamed over a WebSocket and returned as incrementally recognized tokens, not as a single completed transcript. Pin an explicit model, log the version actually resolved, and treat a friendly alias as a moving target. The public API documentation lists a realtime WebSocket service and separately notes regional endpoint domains.
Provisional and final tokens are different products
Each returned token may carry `is_final`. A `false` token is a hypothesis that can change as later context arrives; a `true` token is the finalized form as defined by the provider. Display partial words in a temporary region, then commit final tokens to the archival transcript exactly once. If a network interruption causes duplicate events, use a per-session event identifier or indexed token sequence in your own system and deduplicate before writing captions. A final token is not a human-verified quote, particularly for names, account numbers and overlapping speakers.
What sixty-plus languages really means
Soniox documents a unified supported-language list covering more than 60 languages for realtime and asynchronous recognition. It also documents translation across supported-language pairs, but that does not prove equal recognition quality for every accent or noisy microphone. Retrieve current model language coverage from the vendor endpoint where possible. For Turkish, Arabic, Cantonese or mixed-language customer support, have bilingual annotators mark source audio and score switch boundaries, not only English sentences. Do not force a local script into Latin spelling just to make a default UI easier to render.
Performance measurement that an implementer can repeat
Measure microphone-capture delay, time to first non-final token, time to stable final token and transcription correction effort separately. Also log backpressure, concurrent-session limits, unexpected connection closes and application-region routing. An apparent 150-millisecond token update on a laboratory clip does not guarantee a user’s end-to-end caption delay. Include number reading, overlapping speech, interrupted sentences and quiet utterances in test recordings, and replay identical clips after any SDK or model change.
Security and publishing boundary
The browser should not contain your long-lived organization API secret. Soniox publishes a sample architecture in which a server issues temporary keys for browser or mobile use. Record user notice, audio-retention policy, geographic processing selection and deletion procedure before streaming a call. In regulated or high-stakes work, let a trained reviewer approve transcript citations and any translated captions. A model’s final-state flag is a decoding event, not consent or legal authenticity.
Worked production example
A live support call begins in Polish but the customer quotes an English tracking number, then pauses. Display provisional recognition as subdued text until final tokens arrive and ensure the token renderer does not duplicate the English number after reconnection. Archive the raw packets and final transcript separately; record which temporary hypothesis was shown and when, but never publish the unreviewed screen output as an authenticated transcript.
Practical acceptance gate
Create a six-minute bilingual gold clip containing interruptions and digit strings. Require complete event ordering, zero duplicate committed tokens and documented fallback handling for disconnects. Human-check speech errors for each represented language and measure delay until final tokens, not merely the first provisional syllable. After any alias or API upgrade, rerun exactly the same sample and compare the corrected result.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: availability varies by account/endpoint and jurisdiction. Version: stt-rt-v5 / stt-async-v5. Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Soniox Models — official documentation (checked 2026-10-25)
- Soniox Realtime — official documentation (checked 2026-10-25)
- Soniox Languages — official documentation (checked 2026-10-25)
- Soniox Api — official documentation (checked 2026-10-25)