Soniox STT v6
A speech recognition model released by Soniox that supports bidirectional awareness and echo filtering. It is released in two interface forms: streaming (stt-rt-v6) and asynchronous (stt-async-v6).
Timeline (last 90 days)
Soniox released the STT v6 speech recognition model.
The core mechanism of Soniox STT v6 is to have the model simultaneously receive the caller's speech and the text being spoken by the voice agent as input.
Soniox STT v6 uses the agent text as context to assist in recognizing users' short replies (such as "mm," "right," "no") in order to improve accuracy.
When the agent text does not match what the user actually says (for example, the user contradicts or interrupts), the STT model needs the ability to suppress this text prior, otherwise it may rely on it excessively.
AEC and model-side echo filtering are not entirely mutually exclusive: AEC handles the signal layer and model-side filtering handles the recognition layer, and in actual deployment basic AEC may still be needed to ensure the model input is not overwhelmed by echo.
Echo filtering in streaming scenarios is harder than in asynchronous ones because the model must make judgments before the audio has fully arrived and must not introduce additional latency due to filtering.
At present, comparable evaluation data is still lacking on the actual performance of Soniox STT v6 echo filtering in complex acoustic environments and on the specific magnitude of improvement in short-reply recognition.
All Facts (7)
When the agent text does not match what the user actually says (for example, the user contradicts or interrupts), the STT model needs the ability to suppress this text prior, otherwise it may rely on it excessively.
50%UnverifiedThe core mechanism of Soniox STT v6 is to have the model simultaneously receive the caller's speech and the text being spoken by the voice agent as input.
50%UnverifiedSoniox STT v6 uses the agent text as context to assist in recognizing users' short replies (such as "mm," "right," "no") in order to improve accuracy.
50%UnverifiedAt present, comparable evaluation data is still lacking on the actual performance of Soniox STT v6 echo filtering in complex acoustic environments and on the specific magnitude of improvement in short-reply recognition.
50%UnverifiedAEC and model-side echo filtering are not entirely mutually exclusive: AEC handles the signal layer and model-side filtering handles the recognition layer, and in actual deployment basic AEC may still be needed to ensure the model input is not overwhelmed by echo.
50%UnverifiedEcho filtering in streaming scenarios is harder than in asynchronous ones because the model must make judgments before the audio has fully arrived and must not introduce additional latency due to filtering.
50%UnverifiedSoniox released the STT v6 speech recognition model.
50%