As anyone who as tried transcription engines based on speech recognition, approximately 10% of the resulting accuracy depends on context, which is where all automatic systems fail most spectacularly (automatic closed captions provide the most hilarious and exasperating examples). To further enhance the accuracy of his speech recognition AI and to provide relevant contextual information about the recognized speech in the audio, Palo Alto, California-based Soniox has developed Soniox Knowledge Graph.
The solution has been harnessed from publicly available data and it contains millions of entities that are grouped into hundreds of concepts. The knowledge graph contains structured information about the entities (infoboxes) and the entities themselves are interlinked. Each entity is also linked to the Web with the best-appropriate URL (e.g. Wikipedia page, medical database). The knowledge graph is constantly increasing with more entities and knowledge about the entities. Soniox Knowledge Graph serves as a bridge between speech in the audio and its information on the World Wide Web.
Soniox Knowledge Graph is used in conjunction with Soniox speech recognition AI to automatically transcribe audio and annotate the transcription with recognized entities from the Soniox Knowledge Graph, all happening simultaneously in real-time and low-latency. To support these requirements, Soniox has developed an entity matching engine to efficiently find the entity matches and potentially disambiguate the entities. Soniox offers the recognition and annotation in a single API call, which makes it easy to use and integrate in any downstream applications.
To showcase the new technology, Soniox has integrated Knowledge Augmented Audio AI into Soniox iOS app and Soniox web application. These applications transcribe any audio streams and at the same time highlight the recognized entities in the live transcript. By clicking on the highlighted entity, the additional information about the entity is presented in the form of infobox next to the live transcript of the audio.
"We have developed a new audio augmented experience" says Klemen Simonic, Founder and CEO of Soniox. “We have world-leading speech recognition AI that is now tightly coupled with a large structured knowledge base about the world. Our technology instantly transcribes, annotates and provides relevant information about the speech in the audio. We expect this technology to be used by numerous customers that require actionable insights about the audio.”
The company offers a great demonstration video of its Knowledge Graph contextual associations here.
Soniox was founded in April in 2020 in Redwood City, California, and develops self-learning artificial intelligence for automatic speech recognition. The company invented a novel approach to training speech recognition AI systems, learning to recognize words by exploring different interpretations of spoken words in unlabeled audio and their usage in unlabeled written text. Soniox solution also recognizes speech in different environments, detecting emotions, intonations and spacings between words using sentence boundary or punctuation models based on acoustic information. Soniox is also working on AI models for detection of real-world audio events and deep understanding of audio content.
www.soniox.com
