What is Custom Speech?
With Custom Speech, you can evaluate and improve the Microsoft speech-to-text accuracy for your applications and products.
Out of the box, speech to text utilizes a Universal Language Model as a base model that is trained with Microsoft-owned data and reflects commonly used spoken language. The base model is pre-trained with dialects and phonetics representing a variety of common domains. When you make a speech recognition request, the most recent base model for each supported language is used by default. The base model works very well in most speech recognition scenarios.
A custom model can be used to augment the base model to improve recognition of domain-specific vocabulary specific to the application by providing text data to train the model. It can also be used to improve recognition based for the specific audio conditions of the application by providing audio data with reference transcriptions.
You pay to use Custom Speech models, but you are not charged for training a model. Usage includes hosting of your deployed custom endpoint in addition to using the endpoint for speech-to-text. For more information, see Speech service pricing.
How does it work?
With Custom Speech, you can upload your own data, test and train a custom model, compare accuracy between models, and deploy a model to a custom endpoint.
Here's more information about the sequence of steps shown in the previous diagram:
- Create a project and choose a model. Use a Speech resource that you create in the Azure portal. If you will train a custom model with audio data, choose a Speech resource region with dedicated hardware for training audio data. See footnotes in the regions table for more information.
- Upload test data. Upload test data to evaluate the Microsoft speech-to-text offering for your applications, tools, and products.
- Test recognition quality. Use the Speech Studio to play back uploaded audio and inspect the speech recognition quality of your test data.
- Test model quantitatively. Evaluate and improve the accuracy of the speech-to-text model. The Speech service provides a quantitative word error rate (WER), which you can use to determine if additional training is required.
- Train a model. Provide written transcripts and related text, along with the corresponding audio data. Testing a model before and after training is optional but recommended.
- Deploy a model. Once you're satisfied with the test results, deploy the model to a custom endpoint. With the exception of batch transcription, you must deploy a custom endpoint to use a Custom Speech model.