Inspect Custom Speech data
This page assumes you've read Prepare test data for Custom Speech and have uploaded a dataset for inspection.
Custom Speech provides tools that allow you to visually inspect the recognition quality of a model by comparing audio data with the corresponding recognition result. From the Speech Studio, you can play back uploaded audio and determine if the provided recognition result is correct. This tool helps you inspect quality of Microsoft's baseline speech-to-text model, inspect a trained custom model, or compare transcription by two models.
In this document, you learn how to visually inspect the quality of Microsoft's baseline speech-to-text model, and/or custom models that you've trained . You also learn how to use the online transcription editor to create and refine labeled audio datasets.
Create a test
Follow these instructions to create a test:
- Sign in to the Speech Studio.
- Navigate to Speech-to-text > Custom Speech > [name of project] > Testing.
- Click Add Test.
- Select Inspect quality (Audio-only data). Give the test a name, description, and select your audio dataset.
- Select up to two models that you'd like to test.
- Click Create.
After a test has been successfully created, you can see how a model transcribes the audio dataset you specified, or compare results from two models side by side.
When testing, the system will perform a transcription. This is important to keep in mind, as pricing varies per service offering and subscription level. Always refer to the official Azure Cognitive Services pricing - Speech service for the latest details.
Side-by-side model comparisons
When the test status is Succeeded, click in the test item name to see details of the test. This detail page lists all the utterances in your dataset, and shows the recognition results of the two models you are comparing.
To help inspect the side-by-side comparison, you can toggle various error types including insertion, deletion, and substitution. By listening to the audio and comparing recognition results in each column (showing human-labeled transcription and the results of two speech-to-text models), you can decide which model meets your needs and where improvements are needed.
Side-by-side model testing is useful to validate which speech recognition model is best for an application. For an objective measure of accuracy, requiring transcribed audio, follow the instructions found in Evaluate Accuracy.
Online transcription editor
The online transcription editor allows you to easily work with audio transcriptions in Custom Speech. The main use cases of the editor are as follows:
- You only have audio data, but want to build accurate audio + human-labeled datasets from scratch to use in model training.
- You already have audio + human-labeled datasets, but there are errors or defects in the transcription. The editor allows you to quickly modify the transcriptions to get best training accuracy.
The only requirement to use the transcription editor is to have audio data uploaded (either audio-only, or audio + transcription).
Import datasets to Editor
To import data into the Editor, first navigate to Custom Speech > [Your project] > Editor.
Next, use the following steps to import data.
- Click Import data
- Create a new dataset(s) and give it a description
- Select datasets. Multiple selections is supported, and you can select audio data only, as well as audio + human-labeled data.
- For audio-only data, you can optionally use the default models to automatically generate machine transcription after importing to the Editor
- Click Import
After data has been successfully imported, you can click into the datasets and start editing.
You can also import datasets into the Editor directly by selecting datasets and clicking Export to Editor
Edit transcription by listening to audio
After the data upload has succeeded, click on each item name to see details of the data. You can also use Previous and Next to move between each file.
The detail page lists all the segments in each audio file, and you can click into the desired utterance. For each utterance, you can play back the audio and examine the transcripts, and edit the transcriptions if you find any insertion, deletion, or substitution errors. See the data evaluation how-to for more detail on error types.
After you've made edits, click the Save button.
Export datasets from the Editor
To export datasets back to the Data tab, navigate to the data detail page and click the Export button to export all the files as a new dataset. You can also filter the files by last edited time, audio durations, etc. to partially select the desired files.
The files exported to Data will be used as a brand-new dataset and will not affect any of the existing data/training/testing entities.