What happened
Soniox’s November 3, 2026 announcement introduces STT v6 as the company’s latest speech-recognition model. Soniox says it is available now in both real-time streaming and asynchronous transcription forms. Developers can test v6 on their own audio, use its API, and choose the real-time model for live interactions or the asynchronous model for completed recordings. The announcement does not state a price.
The release centers on two connected features for voice agents. The first, Agent text input, allows developers to stream the agent’s outgoing text into STT v6 as it is sent to text-to-speech. Rather than receiving only the user’s audio, the recognition model can use context from both sides of a dialogue. Soniox says this improves recognition of user replies, including short answers that are ambiguous in isolation, along with names, numbers, specialized terms, and other details shaped by the preceding agent prompt.
The second feature, agent voice echo cancellation, uses Agent text input to distinguish the agent’s playback from user speech. Soniox says STT v6 filters the agent’s own voice out of recognition while continuing to listen to the user during agent speech. That means a user can interrupt an agent without the system treating the agent’s generated audio as user input. Soniox positions the capability as a way to reduce feedback loops and false interruptions on speakerphones, laptops, in-car systems, and other setups where playback can reach the microphone.
Soniox also says v6 improves voice typing, meeting transcription, speaker diarization, spoken-language identification, endpoint detection, and real-time speech-to-text translation. The company describes more than 60 supported languages in one unified model, with language mixing and switching within a sentence supported without developers moving between separate language models. Builders can provide context such as participant names, product names, professional vocabulary, and project terminology. For translation, Soniox says v6 dynamically determines how much speech to gather before translating, seeking enough context for accuracy while keeping output responsive.
What the archive adds
New Runtime’s earlier coverage of Microsoft’s MAI-Transcribe-2-Streaming described speech recognition and speech generation packaged as a stack for conversational agents (October 3, 2026 note). Soniox v6 is a new event in that broader speech-infrastructure thread, with a more specific focus on passing agent-side context into recognition and removing agent playback from the input.
Trend signal
This release supports Agent-ready software exposes capabilities. The recognition layer is not treated as a sealed audio-to-text service: it accepts a defined, task-relevant input from the dialogue system and uses it to interpret the user’s response. It is a vendor-specific implementation, but it illustrates how an agent stack can expose composable, observable capabilities between its speech components.
Why now
Soniox frames the release around a recurring voice-agent failure mode: the agent’s audio is picked up by the microphone, then mistaken for a user utterance. The company also emphasizes multilingual products, where teams need recognition, translation, and language switching without operating separate models for each language.
Why it matters
For a voice-agent team, a transcription error can become an interaction error. It can turn a clear choice into a clarification loop, attach a wrong number to a downstream action, or make the system respond to its own output. Agent text input creates a concrete integration point between the dialogue system, text-to-speech pipeline, and speech-recognition layer: the recognizer can know the question that made a short answer meaningful.
Echo cancellation is equally operationally relevant. Barge-in is a basic expectation for many voice interfaces, but it is difficult when the user and the agent share an acoustic environment. Soniox’s approach, as described, does not require the builder to infer whether a fragment of captured audio belongs to the agent or caller solely from the microphone signal; it uses the known agent text as part of that decision.
The one-model multilingual claim may also reduce architecture complexity for products that handle mixed-language calls or global dictation. But the appropriate decision is not to replace a production transcription stack on the announcement alone. Teams should test whether context improves the fields and conversational turns that matter to their own product, under their actual audio, accents, vocabulary, and latency budget.
Limits and uncertainty
Soniox’s accuracy, speed, and reliability claims are the company’s own. The announcement provides no benchmark scores, test methodology, comparative results, word-error rates, or measurements of the improvement from Agent text input or echo cancellation. It also does not quantify latency during interruption handling, disclose pricing or usage limits, or explain feature differences between real-time and asynchronous operation beyond their stated use cases. Independent evaluation is needed to establish performance across languages, accents, noisy environments, speakerphone configurations, and domain-specific vocabulary.
Skill to build now
Create a small voice-agent evaluation set from real, permissioned interaction patterns: short confirmations, names, numbers, choices following a prompt, and user interruptions while the agent is speaking. Run the same cases with ordinary recognition and with agent text input enabled. Record transcription accuracy, whether the interruption was detected, whether agent audio leaked into the transcript, and end-to-end response latency. This will reveal whether the feature changes a product outcome rather than merely a transcript metric.
Reader takeaway
Soniox STT v6’s most actionable change is not simply a general accuracy claim. It lets a recognition system receive the agent’s own outgoing text and use that context while excluding echoed agent audio. Voice-agent builders should test that combination on their highest-risk turns, especially interruptions and short replies, before treating it as a production advantage.