Speech translation promises instant comprehension, yet a caption that reaches a screen before the sentence ends may need to change its word order, context or speaker reference. Soniox’s 2026 documentation exposes a practical design: source transcript tokens and translated output are related, but not aligned word for word.
One-way and two-way are not interchangeable
The Soniox real-time translation API lets users add a translation block to an audio-streaming request. `one_way` maps detected source speech to a specified target language; `two_way` works with two selected language codes and switches the translation direction. The documentation gives examples using English and Spanish as well as other locale pairs. For a webinar with English speakers and French spectators, use one-way translation. For a help-desk conversation where both sides speak, two-way mode is a better interaction starting point, but it needs careful direction detection and user clarity.
The token stream cannot be used as a timed subtitle master unchanged
Soniox states that recognized source-speech tokens include audio timestamps, while translated tokens do not include the same timestamps; translated chunks are not one-to-one with source words. Therefore copying source timecodes to translated phrases can put a subtitle on screen before the speaker finishes the idea. For archival VTT or SRT, either perform sentence/segment-level bilingual alignment later or assign approximate reading windows after a native-language reviewer adjusts line breaks and timing. Preserve both original and translated outputs with separate timestamps and identifiers.
Check exact supported pairs rather than a large multiplication
The provider advertises more than 60 supported languages and over 3,600 possible language pairs. That arithmetic describes documented language-set permutations, not independent quality validation for every pair. Evaluate a representative matrix: direct English–Turkish, Turkish–English, Arabic–German and common code-switched utterances if those markets matter. Check regional residency routes and whether the selected STT/TTS combination offers compatible languages before promising a speech-to-speech product. Translation and realistic spoken output are separate tasks.
Streaming architecture and interruption
A typical app captures audio, opens a secure WebSocket, streams packets, renders source tokens and then renders translated tokens as they arrive. It must tolerate partial and final transcriptions, source-language changes, late translations and interrupted sockets. Use a bounded on-screen buffer so unstable words do not flash as trustworthy text. Once the live meeting concludes, create a corrected recording-derived transcript and new translation for the permanent record; a low-latency transcript is a useful aid, not necessarily the best permanent subtitle file.
Sensitive meaning and permission
An operator can mistranslate negation, formal titles, medication names or contractual numbers in a way that changes the meaning of a call. For safety-critical domains, route material instructions to qualified bilingual personnel and indicate to participants that automated interpretation is in use. Do not let a translated synthetic voice impersonate the original speaker without a voice-use license. If content will be public, obtain recording and distribution permission independently of translation service access.
Worked production example
A museum streams a French curator presentation with live English captions while a second room displays Turkish captions. Use separate target-language configurations and preserve the original French token sequence for every speaker. When a French sentence rearranges into English, hold enough of the translated phrase to make the subtitle readable instead of assigning each source word a misleading one-to-one timestamp. The live translation may be visible quickly, but a separately edited SRT must be prepared for the permanent recording.
Practical acceptance gate
Review five difficult extracts with professional bilingual editors, including negation, names, prices and pronouns. Validate source and translation streams in both directions where two-way communication is planned. Do not claim that all language pairs perform equally because the service supports their codes; maintain exception pathways for pairs whose quality has not been checked.
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 translation configuration. 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 Translation — official documentation (checked 2026-10-25)
- Soniox Languages — official documentation (checked 2026-10-25)
- Soniox Api — official documentation (checked 2026-10-25)