> For the complete documentation index, see [llms.txt](https://docs.youquantified.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.youquantified.com/devices/voice-emotion.md).

# Voice Emotion

This data source uses a [Speech Emotion Analyzer](https://github.com/tabahi/WebSpeechAnalyzer?tab=readme-ov-file) demo based on the **formantanalyzer** JavaScript library. This algorithm extracts syllables from the Mel-spectrogram of the input and feeds relevant statistical features into a single-layer neural network that classifies. This model was trained on datasets of actors going through dialogues or improvising. It predicts the following categories:

* Happiness
* Sadness
* Anger
* Neutral

From these labels, we output **valence** and **arousal** scores by using the relative happiness probability (for valence) and neutral probability (for arousal).

## Attribution

{% embed url="<https://github.com/tabahi/WebSpeechAnalyzer>" %}