How to use Microsoft’s speech-to-text API to create a dataset, and then train a model with it on the Peltarion platform.
Create a text dataset from voice recordings

Data scientists are always on the lookout for new datasets, but in spite of the myriad of data being generated and processed everyday they can be surprisingly hard to find. Creating your own datasets is therefore often the most time efficient way to go if you want to start exploring a project idea quickly.
Here’s one way you can go about creating a dataset for text using Microsoft’s speech-to-text API, and then using it to train a model on the Peltarion platform.
- Follow this link to find Microsoft’s instructions for their speech-to-text APIs. You can choose to work with recorded audio files or by talking into your microphone.
- Create your dataset by recording people speaking and attaching a label to each recorded snippet.
- Turn it into a text dataset using the speech-to-text API according to Microsoft’s instructions above.
- Follow one of our tutorials to train a text classification model using the natural language processing model English BERT (our tutorial shows you how to build a model to detect the sentiment of reviews) or multilingual BERT (our tutorial on this can tell you the genre of a book based on an extract from it) depending on which languages your recordings are in.
- Deploy the model using our one-click REST API to start using it.
For more about text-based AI models or multilingual BERT check out our intro series on the topic.
- Audio data ,
- Connectors ,
- NLP ,
- Text data ,
Related topics
02/ More on Data science
- Data science /
Towards a token-free future in NLP
- Data science /
Improving multilingual models for low-resource languages
- Data science /
Extending BERT-like Language Models for Multiple Languages and Longer Documents
- Data science /
Predicting Bitcoin prices with AI
You need not have deep pockets or talented engineers to predict Bitcoin prices. Just use Python to download publicly available data and the Peltarion platform to train an AI on it. How hard can it be? As it turns out, it is not hard at all.
- Data science /
Search text by semantic similarity