Soniox STT v6 Released: Adding Text Input and Echo Filtering to Voice Agents

Voice agent interactions have long been plagued by two nagging issues: short user replies are recognized inaccurately, and the agent's own speech gets picked up by the microphone, causing echo interference. Soniox's STT v6 attempts to solve both problems at the model level.

Two-Way Awareness: Letting the STT "Know" What the Agent Is Saying

The core design idea behind Soniox STT v6 is to have the speech recognition model receive two streams of information at once: one is the caller's speech, and the other is the text the voice agent is currently speaking. The model listens to the caller while reading the agent's text, using that text as context to aid recognition.

The first direct benefit of this "two-way awareness" mechanism is the recognition accuracy of short replies. In voice agent scenarios, users often respond with extremely short utterances like "mm," "yeah," or "no," which traditional STT easily misinterprets without context. When the model knows what the agent just asked, the semantic direction of the short reply becomes clear.

In speech recognition research, using external text as a conditioning input to guide decoding is not an entirely new concept; related approaches include contextual biasing and prompt-based shallow fusion: the former is commonly used to inject hotwords such as contact names and proper nouns into the decoding graph to improve recognition, while the latter encodes preceding text into a vector that participates in acoustic model scoring. What makes Soniox special is that what gets injected is not historical transcripts or a hotword list, but the "future utterance" of the other side of the conversation—the agent's reply content that has not yet been heard. This constitutes a form of cross-channel conditioning: while recognizing the caller's speech, the model simultaneously possesses a strongly predictive conversational prior. In theory it can both resolve the semantic ambiguity of short replies and carry the risk of over-reliance—for example, when the agent's text does not match what the user actually says (the user contradicts or interrupts), the model needs the ability to suppress that prior.

Echo Filtering: Removing the Agent's Own Speech at the Recognition Level

Another capability of Soniox STT v6 is using the agent text input to discard the agent's own voice. In scenarios such as hands-free use and speakerphone, the microphone picks up both the user's speech and the agent's playback, creating echo.

Traditional solutions mostly rely on a separate acoustic echo cancellation (AEC) module, whereas Soniox delegates this judgment to the STT model itself—the model already knows the agent's text content and can therefore identify and filter out the corresponding audio component. This approach of embedding echo handling into the recognition stage reduces the voice agent pipeline's dependence on additional audio preprocessing components.

AEC (Acoustic Echo Cancellation) is a classic signal processing module in voice communications. Its principle is to first use an adaptive filter to estimate the acoustic echo path from speaker to microphone, then subtract the predicted echo component from the microphone signal, supplemented by nonlinear residual suppression when necessary. Traditional AEC operates at the audio waveform or spectrum level and does not care what the speech content is, so it is prone to residual echo or damage to the user's voice in situations such as double-talk (user and agent speaking simultaneously), nonlinear distortion, and changing room reverberation. Soniox's approach is equivalent to introducing a "semantic prior" channel: the model knows the text the agent spoke and can directly judge at the recognition stage whether a segment of audio corresponds to this known text, thereby removing it at the text output level. The two are not entirely mutually exclusive—AEC handles the signal layer, while model-side filtering handles the recognition layer, and actual deployments may still require basic AEC to ensure the model input is not overwhelmed by echo.

Two Interface Forms: Streaming and Asynchronous

According to information released by Soniox, STT v6 is released under two model identifiers: stt-rt-v6 for real-time scenarios for streaming recognition, and stt-async-v6 for asynchronous processing. Voice agents are a typical real-time interaction scenario, making the streaming version their primary form; the asynchronous version suits transcription tasks with lower real-time requirements.

This dual-interface division continues the consistent structure of the Soniox API, allowing developers to choose by scenario without having to maintain two tech stacks for different latency requirements.

Streaming STT refers to a working mode in which audio is continuously fed to the model in chunks and recognition results are returned incrementally. It typically requires end-to-end latency within a few hundred milliseconds and supports "partial results" that are revised on the fly as context evolves. Asynchronous/batch STT, by contrast, first uploads the complete audio and then returns a full transcript with timestamps and punctuation all at once; the model can access global context, so accuracy and formatting quality are usually higher, but latency is measured in seconds or even minutes. For voice agents, streaming is essential because the system must quickly determine the user's intent during pauses; for scenarios such as generating meeting minutes or quality-inspection transcripts after a call ends, the asynchronous version is more economical. Notably, echo filtering in streaming scenarios is harder than in asynchronous ones—the model must make judgments before the audio has fully arrived, and it cannot introduce additional latency due to filtering.

What It Means for Voice Agent Developers

The engineering quality of voice agents often gets stuck at the "last mile": the LLM capability is sufficient, but short-reply misrecognition and echo interference at the voice layer significantly degrade the overall experience. Soniox STT v6 consolidates these two problems into the STT model itself, which for developers means less glue code and a simpler audio processing pipeline.

That said, the currently public information is concentrated at the capability description level, and there is still a lack of comparable evaluation data on the actual performance of echo filtering in complex acoustic environments and the specific improvement in short-reply recognition. Whether such capabilities can be reliably delivered in real call scenarios remains to be verified by developers through hands-on testing.

For teams building or optimizing voice agents, the "agent text input" interface offered by Soniox STT v6 is worth attention—it represents a product idea of pushing conversational context forward into the recognition layer, rather than treating STT merely as an independent transcription black box.

One-line Summary

Soniox STT v6 lets the recognition model read agent text, improving short-reply accuracy and enabling recognition-layer echo filtering.

Paragraph Summary

Soniox has released the STT v6 speech recognition model, whose core innovation is "two-way awareness": the model listens to the caller's speech while receiving the text the voice agent is currently speaking, using that text as context to aid recognition. This mechanism delivers two benefits—first, it significantly improves the recognition accuracy of extremely short replies such as "mm," "yeah," and "no," because the model knows what the agent just asked; second, it uses the known agent text to identify and filter the agent's own speech re-captured by the microphone, embedding echo handling from a separate AEC module into the recognition stage. The model is released in two forms: stt-rt-v6 (streaming) and stt-async-v6 (asynchronous), with the former aimed at real-time scenarios such as voice agents. The article also notes that currently public information remains at the capability description level, and there is still a lack of comparable evaluation data on echo filtering performance in complex acoustic environments and the magnitude of improvement in short-reply recognition.