Cross platform CLI to that uses Azure Cognitive Services Speech Service to iterate a directory of audio samples (wav files) and produce a single text file (1 per directory) of all detected speech. The transcript file produced is in the format required for training a Custom Voice Font.
To learn more about Microsoft Cognitive Speech Service, please see here: https://docs.microsoft.com/en-us/azure/cognitive-services/Speech-Service/ and to learn more about Microsoft Cognitive Custom Voice see here: https://docs.microsoft.com/en-us/azure/cognitive-services/speech-service/how-to-customize-voice-font
To retrieve your Speech To Text API key start here: https://azure.microsoft.com/en-gb/try/cognitive-services/
| Argument name | Description | Shortcut | Example |
|---|---|---|---|
| SubscriptionKey | Service Subscription Key | -k | 0a0234adsfasdf |
| Region | Data Center region | -r | Northeurope |
| SourcePath | Directory of files | -p | c:\samples |
The CLI will recursively scan all sub-directories. However, if you are generating transcripts specifically for Custom Voice Font, put all the samples in a single directory to run the CLI:
BulkSSTCLI.exe -k *yourspeechapikey* -r northeurope -p c:\samples
Assuming you have 3 wav files in the c:\samples directory
- 0000000001.wav
- 0000000002.wav
- 0000000003.wav
A new file samples.txt will be created c:\samples\samples.txt containing the recognised speech from each wav file eg:
0000000001 Some recognised speech to text from file 0000000001.wav
0000000002 Some recognised speech to text from file 0000000002.wav
0000000003 Some recognised speech to text from file 0000000003.wav
The format of the text output is ready for use with Custom Voice Font Training
Note, you'll need to ensure your samples are split into separate audio files (each maximum of 60 seconds long), numerically named. You can a split a large audio file easily using ffmpeg eg:
ffmpeg -i "Some.mp3" -f segment -segment_time 59 -ar 16000 -ac 1 %03d.wav
Then run this CLI on the directory containing all the output wav files. This will then generate a single txt file. Take all the wav files, zip them up and you have everything you need to train your own Custom Voice Font.