KeenASR Software Development Kit provides on-device automatic speech recognition functionality for mobile devices running iOS or Android operating system.
Speech recognition is performed on the device; no internet connectivity nor cloud support is required. The SDK is based on the state-of-the-art Deep Neural Network decoder and acoustic models.
The SDK provides Deep Neural Network (DNN) decoders, variety of DNN acoustic models, as well as an API for creating language models on the device for small to medium recognition tasks (up to several thousand words). For large vocabulary dictation-type tasks, language models and decoding graphs can be created ahead of time and used by the SDK. For domain specific large vocabulary use cases, large amounts of domain specific text data might be required to build appropriate langauge models.
The SDK also supports trigger phrase functionality, which is viable for sessions that last up to couple of hours. Current approach for trigger phrase implementaion is not viable for always-on (24x7) listening use cases, especially on battery-powered devices.
Platforms and Development Tools Support
In addition to Objective C iOS framework and Android Java AAR SDK, we also provide instructions on how to use Objective C framework in Swift, as well as Unity plugin which provides C# interface to both iOS and Android versions of the SDK.
iOS and Android SDKs are on separate release paths. Unity plugin is typically updated when there is a new release of either iOS or Android SDK.
Keen Research customers can also leverage Dashboard, a cloud-based tool for on-device ASR development support. KeenASR, in conjunction with Dashboard, provides seamless synchronization of on-device data, and cloud-based access to audio recordings and speech recognition metadata for further analysis and debugging, ability to transcribe and score responses, etc..
Trial Version of the SDK
Keen Research provides trial version of the SDK for both iOS and Android, as well as ASR Bundles (acoustic models) that can be used with the trial SDK.
Currently, we provide English language ASR bundles. ASR Bundles for majority of frequently spoken languages can be provided within 6-8 weeks.
We can also provide customized ASR Bundles that work better in specific acoustic environments, with specific population, and/or have a smaller memory footprint and CPU utilization.