Emotional language

The emotional language model measures 53 dimensions of emotional expression from the meaning and tone of text. It supports 5 additional expressions beyond the other models: Annoyance, Disapproval, Enthusiasm, Gratitude, and Sarcasm. Recommended input filetypes: .txt, .mp3, .wav, .mp4.

You can optionally enable sentiment analysis and toxicity detection alongside emotion scores. The NER model can also be run alongside emotional language to identify named entities in text.

Job configuration

Example job configuration

$
curl -X POST "https://api.hume.ai/v0/batch/jobs" \
>
  -H "X-Hume-Api-Key: <YOUR_API_KEY>" \
>
  -H "Content-Type: application/json" \
>
  -d '{
>
    "models": {
>
      "language": {
>
        "granularity": "sentence",
>
        "sentiment": {},
>
        "toxicity": {}
>
      }
>
    },
>
    "urls": ["https://example.com/audio.mp3"]
>
  }'

Output

Each prediction includes:

• Text: the analyzed text segment
• Position: the begin and end character indices
• Emotion scores: scores for each of the 53 expressions
• Sentiment: distribution over the 9-point scale (when enabled)
• Toxicity: scores for each toxicity category (when enabled)

1
{
2
  "grouped_predictions": [
3
    {
4
      "id": "unknown",
5
      "predictions": [
6
        {
7
          "text": "I'm so happy to see you",
8
          "position": {
9
            "begin": 0,
10
            "end": 23
11
          },
12
          "emotions": [
13
            { "name": "Admiration", "score": 0.107 },
14
            { "name": "Joy", "score": 0.482 },
15
            ...
16
          ],
17
          "sentiment": [
18
            { "name": "1", "score": 0.01 },
19
            ...
20
            { "name": "9", "score": 0.04 }
21
          ],
22
          "toxicity": [
23
            { "name": "toxic", "score": 0.001 },
24
            ...
25
          ]
26
        }
27
      ]
28
    }
29
  ]
30
}

Granularity

The granularity parameter controls how text is segmented before predictions are generated.

Sentiment

When sentiment is enabled, each prediction includes a probability distribution over a 9-point scale, where 1 represents the most negative sentiment and 9 represents the most positive.

1
"sentiment": [
2
  { "name": "1", "score": 0.01 },
3
  { "name": "2", "score": 0.02 },
4
  { "name": "3", "score": 0.05 },
5
  { "name": "4", "score": 0.10 },
6
  { "name": "5", "score": 0.30 },
7
  { "name": "6", "score": 0.25 },
8
  { "name": "7", "score": 0.15 },
9
  { "name": "8", "score": 0.08 },
10
  { "name": "9", "score": 0.04 }
11
]

Toxicity

When toxicity is enabled, each prediction includes scores for the following categories:

Transcription

When processing audio or video with the language model, Hume transcribes speech to text before analysis. Transcription settings are configured separately from models.

$
curl -X POST "https://api.hume.ai/v0/batch/jobs" \
>
  -H "X-Hume-Api-Key: <YOUR_API_KEY>" \
>
  -H "Content-Type: application/json" \
>
  -d '{
>
    "models": {
>
      "language": {
>
        "granularity": "sentence"
>
      }
>
    },
>
    "transcription": {
>
      "language": "en",
>
      "confidence_threshold": 0.5
>
    },
>
    "urls": ["https://example.com/audio.mp3"]
>
  }'

Named Entity Recognition (NER)

The NER model identifies people, places, organizations, and other entities in text. It can be run alongside the emotional language model.

NER accepts one job configuration parameter:

$
curl -X POST "https://api.hume.ai/v0/batch/jobs" \
>
  -H "X-Hume-Api-Key: <YOUR_API_KEY>" \
>
  -H "Content-Type: application/json" \
>
  -d '{
>
    "models": {
>
      "language": {
>
        "granularity": "sentence"
>
      },
>
      "ner": {
>
        "identify_speakers": true
>
      }
>
    },
>
    "urls": ["https://example.com/audio.mp3"]
>
  }'

Expressions

The emotional language model measures the following 53 expressions. The 5 expressions marked with * are unique to the language model and not available in the face, prosody, or vocal burst models.