For years, companies have tried and failed to build voice-powered products that work in the real world. Support calls mishear customer names or addresses. Medical transcripts miss key details and drop important acronyms. Meeting notes mix up speakers or simply merge them into one.
Even with today’s models, speech-to-text still can’t handle what real life sounds like. People mumble, pause, and talk over each other. They say numbers, codes, and brand names that don’t exist in a dictionary. They speak with accents and mix phrases from different languages. The moment speech leaves a studio setting, accuracy leaves with it.
To make up for it, companies chain APIs, write cleanup scripts, or add humans back into the loop. Devs waste time debugging words, instead of building new products and experiences. And the result is always the same: complexity, cost, and unreliable data at scale. Speech AI never truly understood speech – until now.
A new standard for understanding speech
Our latest version of Soniox v3 is setting a new standard for speech AI. It delivers breakthrough accuracy, faster and more precise language detection, and higher quality translation – all powered by a single foundation model that understands real-world speech in deep context. It’s like having native speaker accuracy in 60+ languages.
With v3, every part of the system has evolved:
- Breakthrough accuracy that holds up in natural, fast, and overlapping speech for every language
- Instant, reliable language detection that recognizes and switches languages seamlessly
- Smarter translation that captures meaning, not just literal phrasing
- High accuracy on alphanumerics, such as phone numbers, addresses, and IDs, for every language
- Built-in domain intelligence that understands names, acronyms, and context across industries and topics
- More capacity and scale for longer recordings, more users, and real-time streaming at once
Soniox v3 captures meaning, not just words. It knows when someone is spelling out an email address, giving a serial code, or describing a customer issue, and gets it right in real time.
“It just gets the words right — any language, any accent, any context. That’s what accuracy is supposed to look like.”
“Soniox has solved this problem with exceptional recognition of alphanumeric sequences, resulting in a much higher acceptance rate for our voicebot.”
“It just gets the context — and when we add our own domain knowledge, it feels completely customized to us.”
Build better voice experiences with Soniox v3
Soniox v3 opens the door to real-world voice products that were never reliable before: multilingual customer support bots, accurate meeting assistants, real-time translation tools, and medical, legal, or technical transcription that’s instantly usable.
Soniox v3 is available in both real-time and asynchronous modes, with two models being released: stt-rt-v3 and stt-async-v3.
Start transcribing, translating, and understanding speech with the Soniox API. Soniox v3 is available today through our API and on our mobile app.