Mozilla DeepSpeech has been updated with support for TensorFlow Lite, resulting in a smaller package size and faster performance on some platforms.
DeepSpeech is a deep leaning-based automatic speech recognition (ASR) engine with a simple API developed by Mozilla. The speech recognition technology and trained models in DeepSpeech are openly available to developers. Mozilla also provides pre-trained English models.
The latest release, version v0.6, comes with support for TensorFlow Lite, the version of TensorFlow that’s optimized for mobile and embedded devices. This has reduced the DeepSpeech package size from 98 MB to 3.7 MB, and cut the English model size from 188 MB to 47 MB. The developers achieved the cut using post-training quantization, a technique to compress model weights after training is done.
While TensorFlow Lite is designed for mobile and embedded devices, the DeepSpeech team found it also made DeepSpeech faster on desktop platforms, so they’ve made it available on Windows, macOS, and Linux as well as Raspberry Pi and Android. DeepSpeech v0.6 with TensorFlow Lite runs faster than real time on a single core of a Raspberry Pi 4. It also uses 22 times less memory.
The move to TensorFlow 1.14 means the developers have been able to make use of the CuDNN RNN APIs for their training code. This change has given them around two times faster training times, which means faster experimentation and better models. Support for online feature augmentation has also been added.
Along with the performance improvements, the new decoder exposes timing and confidence metadata, providing new possibilities for applications. The new release also includes an extended set of functions in the API, not just the textual transcript. You also get metadata timing information for each character in the transcript, and a per-sentence confidence value.