Train a model for Custom Speech
Training a speech-to-text is necessary to improve recognition accuracy for both Microsoft's baseline model or a custom model that you're planning to create. A model is trained using human-labeled transcriptions and related text. These datasets along with previously uploaded audio data, are used to refine and train the speech-to-text model to recognize words, phrases, acronyms, names, and other product-specific terms. The more in-domain datasets that you provide (data that is related to what users will say and what you expect to recognize), the more accurate your model will be, which results in improved recognition. Keep in mind, that by feeding unrelated data into your training, you can reduce or hurt the accuracy of your model.
Use training to resolve accuracy issues
If you're encountering recognition issues with your model, using human-labeled transcripts and related data for additional training can help to improve accuracy. Use this table to determine which dataset to use to address your issue(s):
|Use case||Data type||Data quantity|
|Proper names are misrecognized||Related text (sentences/utterances)||10 MB to 500 MB|
|Words are misrecognized because of an accent||Related text (pronunciation)||Provide the misrecognized words|
|Common words are deleted or misrecognized||Audio + human-labeled transcripts||10 to 1,000 transcription hours|
If you haven't uploaded a data set, please see Prepare and test your data. This document provides instructions for uploading data, and guidelines for creating high-quality datasets.
Train and evaluate a model
The first step to train a model is to upload training data. Use Prepare and test your data for step-by-step instructions to prepare human-labeled transcriptions and related text (utterances and pronunciations). After you've uploaded training data, follow these instructions to start training your model:
- Navigate to Speech-to-text > Custom Speech > Training.
- Click Train model.
- Next, give your training a Name and Description.
- From the Scenario and Baseline model drop-down menu, select the scenario that best fits your domain. If you're unsure of which scenario to choose, select General. The baseline model is the starting point for training. If you don't have a preference, you can use the latest.
- From the Select training data page, choose one or multiple audio + human-labeled transcription datasets that you'd like to use for training.
- Once the training is complete, you can choose to perform accuracy testing on the newly trained model. This step is optional.
- Select Create to build your custom model.
The Training table displays a new entry that corresponds to this newly created model. The table also displays the status: Processing, Succeeded, Failed.
Evaluate the accuracy of a trained model
You can inspect the data and evaluate model accuracy using these documents:
If you chose to test accuracy, it's important to select an acoustic dataset that's different from the one you used with your model to get a realistic sense of the model’s performance.