Training a device specific acoustic model
Developers can configure custom acoustic models by providing audio files with custom recordings. The custom acoustic model is trained by tuning a baseline model using the audio files. Audio files may contain custom noise to apply to clean speech for training. The custom acoustic model is provided as an alternative to a standard acoustic model. Device developers can select an acoustic model by a user interface. Speech recognition is performed on speech audio using one or more acoustic models. The result can be provided to developers through the user interface, and an error rate can be computed and also provided.
1. A method comprising:
providing a user interface for developers to select between a standard acoustic model and a custom acoustic model;
receiving, from a developer, one or more audio files of custom recordings appropriate for a specific device;
training the custom acoustic model from a baseline acoustic model tuned to the custom recordings;
receiving, from the developer, a selection of the custom acoustic model; and
deploying the custom acoustic model for speech recognition by the specific device.
2. The method of claim 1 wherein the one or more audio files of custom recordings contain custom noise and the baseline acoustic model is based on clean speech.
3. The method of claim 1 further comprising:
performing automatic speech recognition of speech in the one or more audio files using the standard acoustic model to determine a first transcription;
performing automatic speech recognition of speech in the one or more audio files using the custom acoustic model to determine a second transcription; and
providing an output to a visual display showing differences between the first transcription and the second transcription.
4. The method of claim 1 further comprising:
receiving, from the developer, labeled transcription data corresponding to the one or more audio files;
performing automatic speech recognition of speech in the one or more audio files using the custom acoustic model to determine machine transcription data;
comparing the labeled transcription data to the machine transcription data to compute an error rate; and
providing the error rate to a developer.