Soniox adds agent speech context to its v6 recognition models Soniox adds agent speech context to its v6 recognition models

Context from the agent

Soniox lets developers supply the agent’s outgoing words alongside microphone audio, giving recognition access to that conversational context. The company also says the model can exclude the agent’s playback from the transcript while continuing to recognize a caller who interrupts. Existing acoustic echo cancellation should remain enabled. These two features are available in the real-time model across supported languages. An asynchronous model handles recordings. The launch also describes improvements to dictation, translation and speaker identification, without supplying a comparative accuracy result in the announcement.

Playback controls the text stream

The dated JavaScript client makes timing part of the interface. Soniox places each text message among the incoming audio frames. Developers send words shortly before they play, rather than forwarding a whole response as soon as the language model generates it. Speech synthesis can also run ahead of playback. The client’s documentation specifies a lead of 500 milliseconds to one second, paced against how much audio has actually played. At an interruption, the application sends any already-played text still outstanding, marks the segment finished and discards the unspoken remainder. Each segment represents an agent turn, not an individual sentence. Spaces must remain intact because the service joins text pieces exactly as supplied. The core session method rejects a send if agent text input was not enabled in its configuration. It also rejects calls while the session is disconnected. Applications therefore need to handle client state as well as scheduling the text against playback.

Compatible requests, different written output

The model changelog lists changes that matter beyond recognition quality. Currency amounts retain spoken currency words instead of automatically becoming a symbol-and-decimal representation. Alphanumeric identifiers receive hyphens only when those hyphens were spoken. Dictation commands for punctuation or editing remain words, leaving the application to apply them. A parser that expects a normalized amount or a particular identifier layout therefore needs more than a successful request as its acceptance test. Soniox says v6 uses the existing interface: an application can switch by changing the requested model name. The old real-time and asynchronous v5 models are scheduled for removal on December 15. After that date, requests naming v5 automatically route to the corresponding v6 model. Keeping the old name will consequently not preserve the old transcription behavior. The transition is an output-contract change even where request syntax remains compatible.