---
title: "Soniox releases STT v6 to help voice agents listen while they speak"
description: "Soniox STT v6 is live for real-time and async transcription. Explore agent text input, echo filtering, benchmark caveats and the December v5 transition."
url: "https://promptszone.com/soniox-stt-v6-release-agent-text-echo-filtering"
published: "2026-11-05T09:00:05Z"
categories: ["New Models"]
---

# Soniox releases STT v6 to help voice agents listen while they speak

Soniox’s new speech models are active, with real-time text context and echo filtering. Benchmark and output-format caveats matter before migration.

![Original diagram showing agent text and microphone audio feeding Soniox STT v6 recognition to produce a user transcript with agent playback filtered out.](https://promptszone.com/assets/uploads/684177f812d991367642a45ff7b0c244.png)

**Summary:**

STT v6 adds context from an agent’s outgoing text and filters its playback. The real-time and async models are live, while comparison data needs careful handling.

Soniox has released STT v6, a speech-recognition update designed to help voice agents understand a conversation while avoiding a familiar failure: mistaking their own spoken replies for the user. The [November 3 announcement](https://soniox.com/blog/soniox-stt-v6) is accompanied by active real-time and asynchronous API models, rather than a waitlist or a promise of later access.

The practical change is that a developer can send an agent’s outgoing text alongside microphone audio. Soniox says that context helps the model interpret short replies and filter the agent’s playback while continuing to recognize a user who interrupts. Those capabilities are worth testing separately from general transcription accuracy.

## What is available now

Soniox’s [model documentation](https://docs.soniox.com/stt/models) lists **stt-rt-v6** for real-time streaming and **stt-async-v6** for recorded-audio processing as active. The company describes improvements across more than 60 supported languages, including transcription, speaker diarization, language identification and translation. These improvement claims come from Soniox; they are not measurements from a PromptsZone hands-on test.

The new agent text input and agent voice echo cancellation features specifically belong to the real-time model. Developers enable agent text input and send what their agent is saying over the same WebSocket used for incoming audio. The asynchronous model’s availability should not be read as support for that live conversational workflow.

## Why the agent’s own words matter

A short answer can be difficult to transcribe without the question that prompted it. An assistant offering two unfamiliar place names, for example, supplies useful context for recognizing the user’s choice. Agent text input gives the recognizer access to that missing side of the exchange.

Playback creates another problem: a microphone may capture both a person and the synthetic voice coming from a speaker. Soniox says its text-aware filtering can remove the agent’s voice from the transcript without stopping recognition when the person talks over it. The [implementation guidance](https://docs.soniox.com/stt/rt/agent-text-input) says to keep acoustic echo cancellation enabled. The model-level feature handles what that earlier audio-processing stage leaves behind; it is not a reason to disable it.

Timing matters: the documentation says to send text roughly 500 milliseconds to one second before the corresponding audio is played, following playback rather than dumping a whole generated response into the stream. When a user interrupts, the application should end the agent’s text segment and stop sending the unplayed remainder. Agent text is billed as input text tokens.

## Why this release does not yet have a clean comparison scorecard

Soniox’s [own benchmark page](https://soniox.com/benchmarks) still identifies its evaluated model as stt-rt-v5, with an August 2026 evaluation. Those results should not be relabeled as v6 evidence.

Independent v6 results do exist on [Speko’s speech-to-text leaderboard](https://benchmarks.speko.ai/). However, the displayed board and its linked [data export](https://benchmarks.speko.ai/data.json) return different v6 figures. The export separates read-English and conversational tests and carries metric-specific measurement dates; we could not reconcile it with the displayed scores. We therefore do not combine those figures into a ranked comparison or claim that v6 leads its peers.

General word-error and latency results would also not, on their own, establish the benefit of the new agent-text and echo-filtering features. That requires a test with those features and the audio conditions explicitly controlled.

## Migration needs output checks, not just a new model name

Soniox says v6 is compatible with the existing API and that developers can switch by changing the requested model name. It also documents output changes that could matter to applications: spoken prices retain currency words, and hyphens in identifiers appear only when spoken. Dictation commands such as “new line” are written as words for the application to interpret.

That makes a regression test especially useful for systems that extract amounts, account references or structured fields from transcripts. API compatibility does not guarantee that downstream formatting assumptions will remain valid.

The company plans to remove stt-rt-v5 and stt-async-v5 on **December 15, 2026**. Its documentation says requests using those names will automatically route to their v6 replacements afterward. Teams should test the explicit v6 versions before relying on that transition.

## What to measure in a pilot

A useful evaluation would pair ordinary transcription checks with conversations recorded on the actual target devices. Test user interruptions during agent playback, brief answers containing unfamiliar names, code-switching and spoken numbers. Measure false interruptions and missed user speech as well as word errors and tail latency. For an echo-filtering comparison, hold the microphone, speaker volume, acoustic processing and agent text feed constant.

The release gives developers a concrete new way to supply conversational context. Whether it improves a particular voice product depends on that product’s audio environment and failure costs, not simply on a broad accuracy claim.

*Image: original PromptsZone explanatory diagram of Soniox’s documented real-time workflow, not a product screenshot or measured result. The unchanged official Soniox STT product-family logo is used under [Soniox’s brand guidelines](https://soniox.com/brand) as the model-logo fallback; no endorsement is implied. Availability and documentation checked November 5, 2026.*