STT 
Today, we’re introducing Soniox STT v6, the world’s first speech recognition model that knows what your voice agent is saying and automatically filters out the agent’s own voice.
Two connected innovations make this possible. Agent text input gives recognition context from both sides of the conversation, improving its understanding of user responses. Agent voice echo cancellation keeps the agent’s playback out of recognition while continuing to listen to the user, even when they interrupt.
Alongside these innovations, v6 delivers significantly higher accuracy across 60+ languages, with improvements in voice typing, meeting transcription, multilingual recognition, and real-time translation. It also advances speaker diarization, language identification, endpoint detection, and context customization.
Available in both real-time and asynchronous models, Soniox STT v6 gives developers a stronger foundation for voice agents, voice typing, and AI meeting products built for a global audience.
Your speech model should know what your agent says
What the agent says shapes what the user says next. A question, a list of options, or a request to confirm a detail provides valuable context for recognizing the response.
Yet speech recognition models typically receive only the user’s audio, leaving them unaware of what the agent just said. STT hears only one side of the conversation.
With Agent text input, developers can stream the agent’s outgoing text into Soniox STT v6 as it is sent to text-to-speech, giving v6 context from both sides of the conversation.
V6 uses this context to recognize user responses significantly more accurately, including names, numbers, specialized terms, and short answers that are often ambiguous in isolation.
More accurate recognition means fewer clarification loops, less frustration, and conversations that move forward smoothly.
Your agent should never interrupt itself
When an agent’s voice is picked up by the microphone, speech recognition can mistake it for user speech. The agent may interrupt itself, respond to its own words, or enter a feedback loop that breaks the conversation.
Soniox STT v6 introduces agent voice echo cancellation. Using Agent text input, v6 distinguishes the agent’s playback from user speech and automatically filters it out of recognition.
Crucially, v6 keeps listening to the user while the agent speaks. When the user interrupts, the model recognizes their speech while filtering out the agent’s voice.
This means fewer false interruptions and more reliable conversations across speakerphones, laptops, in-car systems, and other devices where playback reaches the microphone.
The most accurate, fastest voice typing yet
Soniox STT v6 delivers our fastest, most accurate voice typing across 60+ languages, giving developers the foundation to build reliable applications where voice is the input.
Voice typing demands precision. A single wrong digit changes an amount. A missing unit changes a measurement. A misrecognized name or technical term forces the user to stop and correct the text, breaking the natural flow that makes voice typing useful in the first place.
For v6, we invested heavily in real-world dictation data, with training and evaluation spanning diverse applications, specialized vocabulary, speaking styles, and languages.
The result is significantly more accurate recognition and written rendering of numbers, currencies, units, dates, alphanumeric identifiers, names, and specialized terminology. Developers can further improve accuracy by supplying context relevant to the user’s task, including product names, professional vocabulary, and project terminology.
Speed matters just as much as precision. V6 streams text while the user speaks and finalizes promptly when they finish, keeping voice input responsive throughout the writing workflow.
Build voice typing into editors, messaging apps, forms, medical software, and enterprise tools, with the same speech capabilities across all 60+ supported languages.
The strongest foundation for AI meeting products
Soniox STT v6 redefines meeting transcription and speaker diarization, powering AI notetaking, meeting assistants, and conversation intelligence across 60+ languages. Getting both what was said and who said it right is essential: a missed name, incorrect number, or misattributed statement can distort summaries, decisions, and action items.
Behind these advances is an extensive investment in training data that captures the complexity of real meetings: rapid speaker changes, interruptions, overlapping voices, heated discussions, similar-sounding speakers, and varied microphone conditions. This work substantially improves transcription and speaker diarization as participants exchange ideas, challenge each other, and switch languages.
Context makes recognition even more powerful. Supply participant names, company terminology, product names, and project vocabulary to give v6 knowledge it could not reliably infer from audio alone.
Use the real-time model for live meetings or the asynchronous model for completed recordings. Both provide a more accurate, speaker-aware record for downstream AI.
Bring the full voice experience to 60+ languages
V6 raises recognition accuracy across all 60+ languages, building on the native-speaker accuracy of v5 with better handling of accents, multilingual conversations, and language switching within the same sentence.
The improvements extend across all core features: more accurate speaker diarization, spoken language identification, and endpoint detection, and much stronger use of supplied context. Agent text input and agent voice echo cancellation are available in every supported language, giving developers the new agent capabilities across their global products.
All of this runs through one unified speech AI model. Languages can be mixed freely while preserving the full feature set, without managing or switching between separate language models. Every supported language gets the full feature set available in its real-time or asynchronous model.
More accurate translation, delivered in real time
Soniox STT v6 delivers more accurate real-time speech-to-text translation across 60+ languages, with better preservation of meaning and greater precision for dates, numbers, currencies, and other critical details.
V6 dynamically adjusts how much speech it gathers before translating, capturing the context needed for an accurate result while keeping output responsive. This is especially important in languages such as German and Korean, where information arriving later in a phrase can determine how earlier words should be translated.
V6 also improves multilingual speech translation, including conversations that switch languages mid-sentence. More accurate recognition and translation of alphanumerics help preserve amounts, dates, and identifiers from the original speech through to translated text.
Start building with Soniox STT v6
Available now for real-time streaming and asynchronous transcription. Test v6 on your own audio and experience the improvements firsthand.
Try it live or explore the API documentation to start building.