linnet/docs/user-guide/tone-annotation.md

2.6 KiB

Tone Annotation

Tone annotation labels each speech segment with its emotional and social register — not just what was said, but how it was said.

Why it matters

Reading tone is cognitively expensive. For autistic and neurodivergent (ND) users, parsing the social subtext of speech in real time can compete with comprehending the content itself. Tone annotation offloads that work: Linnet surfaces the register so you can stay focused on what is being communicated rather than spending bandwidth on how it is being delivered.

This is especially useful in high-stakes conversations where the stakes of misreading tone are high: medical appointments, workplace meetings, negotiations, difficult family calls.

How it works

Linnet runs a local SER (speech emotion recognition) model — a fine-tuned wav2vec2 checkpoint — against each audio segment as it is transcribed. No audio leaves your machine. The model outputs a register classification, which Linnet maps to a human-readable label.

Example tone labels

Label What it signals
Apologetic The speaker is hedging or expressing regret
Stressed Elevated pitch or pace suggesting pressure
Enthusiastically agreeing Affirmation with genuine energy
Passive-aggressively complying Verbal agreement with tonal resistance
Dismissive Low engagement, clipped delivery
Reassuring Calm, warm, stabilizing register
Neutral No strong register detected
Uncertain Halting, questioning delivery
Firm Definitive, little hedging
Frustrated Compressed, effortful delivery

The label set is not exhaustive — register is continuous and context-dependent. The model assigns the closest match from the taxonomy.

Confidence and edge cases

Low-confidence segments (where the model is uncertain about the register) are shown with a lighter label. If a segment has no clear tonal signal, it is labeled Neutral.

The model performs best on:

  • Clear audio with minimal background noise
  • English speech (other languages reduce accuracy)
  • Single-speaker segments (diarization helps isolate voices)

Cross-cultural register differences can affect label accuracy — some expressions of politeness read as passive-aggressive to a model trained primarily on Western English corpora. Use tone labels as a signal, not a verdict.

Viewing tone in context

Tone labels appear as tags next to each transcript line in the default view. Switch to Elcor mode to show tone as bracketed prefixes inline with the text.

The full tone timeline is available in the session detail view and is included in Meeting Notes exports.