Soniox Releases STT v6 Speech Model That Listens to What Voice Agents Say

The release comes in two models, stt-rt-v6 for real-time streaming and stt-async-v6 for recorded files, and covers more than 60 languages. It is a hosted service available only through Soniox's cloud API. No self-hosted or on-device version has been announced.

How It Works

In a typical voice agent, the speech recognition layer hears only the caller and never sees the question the caller is answering. That makes short replies easy to get wrong. Soniox's documentation gives examples such as "River Roast" being transcribed as "River Road". A second problem is echo: the agent's own voice leaks back through speakerphones and laptops, gets transcribed as if the user said it, and the agent ends up interrupting itself.

With v6, developers switch on agent text input with a single setting and stream the agent's words over the same connection as the caller's audio. The feature works only with the streaming model. Soniox recommends sending each piece of text 500 milliseconds to one second before it is played. The agent's text does not appear in the transcript the model returns.

Soniox says the model also uses that text to separate leftover agent audio from the caller's speech and filter it out, while still listening so callers can interrupt. MarkTechPost notes this is meant to handle what conventional acoustic echo cancellation misses, not to replace it.

Other changes include better handling of numbers, currencies, dates, IDs and specialist terms, improved meeting transcription and speaker identification, and stronger performance on accents and on speakers switching languages mid-sentence. One formatting change may affect existing apps: v6 writes out amounts as spoken, such as "15 dollars and 20 cents" rather than "$15.20".

Pricing and Claims

Soniox lists pricing of about $0.12 an hour for streaming and $0.10 an hour for file transcription, with speaker identification and language detection included. Agent text is billed as input text tokens at $4.00 per million on the real-time model. MarkTechPost's comparison puts streaming prices from AssemblyAI, Deepgram and ElevenLabs at between $0.39 and $0.55 an hour.

The company describes v6 as "significantly" more accurate than v5 but has not yet published benchmark results for it. Its v5 model scored a 1.34% semantic word error rate on the open-source Pipecat benchmark. Soniox also calls v6 the first model that knows what a voice agent is saying. MarkTechPost points out that AssemblyAI added a similar agent-context feature to its Universal-3.5 Pro Realtime model in July 2026, though it found no mention of echo filtering in AssemblyAI's documentation.

The v5 models are deprecated and will be removed on December 15, 2026, after which requests will be routed to v6 automatically. That gives existing customers about six weeks to test the new formatting behaviour.

For companies running AI phone agents in customer service, transcription errors on short answers and self-interruptions are among the most common causes of failed calls. Feeding the agent's script into the recogniser tackles both problems, but buyers will have to wait for independent v6 benchmarks to judge how much it helps.