Chitti app logo

Chitti is better in the app

Coins, streaks and builds — free on Android & iOS

Get the app
Chitti logoChitti

Voice Recognizer

5 Minutes · Module 5 - Audio Processing

Voice recognition in Teachable Machine teaches students how to train an audio model to identify different voices. A Tamil technology lesson.

  • Differences between image models and audio models
  • Recording background noise and voice samples using a mic
  • Training the audio model and testing live voice predictions
What this lesson teachesRead the lesson as text

A voice recognizer is an audio machine learning model that learns to detect different sounds and identify who is speaking. It works by analyzing audio samples recorded through a microphone and calculating the probability that a sound belongs to a particular person or to ambient background noise.

Audio models versus image models

Training an audio model in Teachable Machine differs from training an image model in two main ways. Image projects allow you to add visual classes immediately, but audio projects require you to record background noise before adding voice classes. Furthermore, while image models accept uploaded picture files like JPG and PNG easily, audio projects rely primarily on recording live sound through a microphone rather than uploading MP3 files.

Collecting audio samples

  • The first step is recording twenty seconds of room sound using a microphone to establish a background noise baseline.
  • Clicking the extract sample button breaks the recorded audio into usable training samples for the model.
  • A new class is renamed for the first speaker, who records multiple voice clips speaking short phrases into the microphone.
  • Another class is created for a second speaker, recording different words and greetings like hello or vanakkam (வணக்கம்) to help the model tell voices apart.

Training and real-time testing

After gathering samples for background noise and each speaker, clicking the train model button allows the system to process the audio data. Once training is complete, the model tests live microphone sound by calculating prediction possibilities shown as percentages. When the room is quiet, the background noise class shows the highest percentage. When a specific person speaks into the microphone, their corresponding class percentage increases significantly.

Saving the finished project

Once testing confirms that the voice recognition model accurately identifies each speaker and the background noise, the project can be downloaded and saved to your computer as a file. This allows you to store the trained model, similar to how voice-activated computers and phones in movies unlock only for a specific person.

Part of the course

A.I. with Teachable Machine

View course →