Speech Technology Magazine

AudioShake, an audio separation technology provider, today launched The Refinery, a system that turns raw, real-world audio into structured, training-ready data at scale. AudioShake's The Refinery can separate people speaking simultaneously into individual speaker tracks, while also isolating dialogue, music, and background sound from finished recordings.

Underheard, a Swiss startup that officially launched today as a spinoff from SFF Studio, the venture studio of Swiss Founders Fund, has released an artificial intelligence interviewer platform that conducts customer interviews for qualitative feedback at any hour and turns the answers into structured findings. Underheard's AI interviewer uses voice AI specifically for qualitative research.

Sightview, an electronic health record and practice management partner focused solely on eyecare, has launched Sightview AI Scribe within its ophthalmology EMR/PM platform. AI Scribe was trained on ophthalmology-specific terminology and speech patterns to recognize provider language and input notes into the correct fields in the chart. The scribe leverages the user's typical code sets and documentation habits to turn words into action.

Voiskey, a voice-typing application provider, has launched its Team feature on its desktop application for Mac and Windows as part of its broader Team Expression Solution. The update helps teams improve expression efficiency by building a shared communication foundation for clearer, faster, and more consistent workplace collaboration. Nowhere is the demand for expression efficiency higher than in the workplace, where clarity and speed directly drive outcomes.

Haley Hilt might not speak, but her grandfather, Ed Fennell, will tell you she says plenty. After nearly two decades caring for Haley, who has a rare neurodevelopmental disorder, Fennell recognizes the subtle gestures, expressions, and sounds that signal when she is happy, uncomfortable, tired, or simply wants something. But he also worries about the future: Who will understand those signals when the people who know her best are no longer there?

Voice search often appears to fail at the last few words. A person can speak a perfectly ordinary request—find a service near a neighborhood, ask whether a business covers a town, or look for a provider on the Cape—and receive a result from the wrong place. The speech recognition might be accurate. The system heard every word. What failed was the local meaning. Location modifiers are deceptively difficult because people use geography differently from databases.

Wordly, a provider of translation and captioning technology, today introduced Silent Theater integration to help trade show exhibitors deliver silent-theater audio through an artificial intelligence translation platform. The new capability streams booth audio directly to attendees' phones with conference-grade clarity and lower latency.

LEIFRAS, a sports and social business company, has partnered with PST, a developer of voice biomarker technologies, to integrate PST's voice analysis platform, VOISLOG into its nationwide sports instruction platform. VOISLOG analyzes voice characteristics to visualize changes in an individual's physical and mental condition, offering a potential tool for well-being support.

Klang is releasing Pianissimo, a speech-to-text model trained for Swedish as open source, freely available for others to use and build on. Pianissimo builds on Nvidia Parakeet, an open speech-to-text model for transcription, which Klang has optimized for Swedish and Swedish dialects. Klang's transcription model reaches around 3,600 times real time, which means it can process roughly an hour of audio in one second. That speed makes it possible to transcribe speech continuously during a conversation.

One of the most dangerous speech-to-text failures is also one of the easiest to miss. The transcript reads cleanly. Then you inspect the value the workflow actually needs: DHCP became DHP, an IP address lost an octet, or a spelled command was normalized into an ordinary word. Unlike a transcription mistake that messes up "thinking" and "tinkering," mistakes of this critical value cannot be recovered later by a large language model in the voice agent system.