| # MELD English Test Sets |
|
|
| ## Emotion Label Alignment |
|
|
| | UsTwo Pipeline (EN) | MELD Label | Match | |
| |---|---|---| |
| | neutral | neutral | ✅ Exact | |
| | joy | joy | ✅ Exact | |
| | sadness | sadness | ✅ Exact | |
| | anger | anger | ✅ Exact | |
| | surprise | surprise | ✅ Exact | |
| | fear | fear | ✅ Exact | |
| | disgust | disgust | ✅ Exact | |
|
|
| **7/7 labels match exactly.** No mapping needed. |
|
|
| ## Test Sets |
|
|
| | File | Scenario | Speakers | Primary Emotion | Duration | Utterances | Emotion Distribution | |
| |---|---|---|---|---|---|---| |
| | 01_angry_fight | Couple in a heated argument | Ross, Rachel | anger | 36.9s | 11 utts | anger:7 neutral:2 sadness:1 disgust:1 | |
| | 02_happy_loving | Couple being affectionate and playful | Chandler, Monica | joy | 43.8s | 16 utts | joy:6 surprise:5 anger:3 neutral:1 sadness:1 | |
| | 03_sad_emotional | Emotional confession — "you still love me?" | Ross, Rachel | sadness | 55.4s | 17 utts | neutral:7 sadness:4 anger:3 surprise:3 | |
| | 04_surprise_shock | Drunk voicemail surprise scene | Ross, Rachel | surprise | 30.2s | 8 utts | surprise:5 neutral:2 sadness:1 | |
| | 05_fear_anxiety | Anxious and worried conversation | Chandler, Rachel | fear | 19.9s | 12 utts | fear:5 neutral:4 surprise:2 sadness:1 | |
| | 06_disgust_annoyance | Annoyed and frustrated bickering | Joey, Rachel | anger | 29.2s | 11 utts | anger:5 neutral:2 sadness:1 surprise:1 fear:1 joy:1 | |
| | 07_bittersweet | Saying goodbye with conflicting feelings | Ross, Rachel | sadness | 43.3s | 14 utts | sadness:6 surprise:3 anger:3 fear:1 neutral:1 | |
| | 08_calm_daily | Casual everyday chitchat (baseline) | Joey, Monica | neutral | 40.4s | 15 utts | neutral:13 joy:2 | |
| | 09_opposite_emotions | Tense speaker + calm listener — triggers `listening` pair animation | 2 spk | surprise | 24.9s | 11 segs | surprise:5 neutral:3 anger:1 fear:1 joy:1 (pipeline fused) | |
| |
| ## Notes |
| - All dialogues are 2-speaker (male + female) conversations |
| - 03: Ross+Rachel only, utt15 removed (timestamp overlap with utt14) |
| - 04: starts from "Rach, I got a message from you", utt3/utt8 removed (timestamp overlaps) |
| - 06: Joey+Rachel only, utt16 removed (addresses Ross) |
| - 07: utt11 removed (timestamp overlap with utt10) |
| - 09: **opposite-emotion demo scene (updated 2026-04-22)** — WAV replaced by user. E2E pipeline output: speaker_0 dominant=surprise (max 0.68, tense via anger residual), speaker_1 dominant=neutral (max 0.54, calm). `(tense, calm)` → **pair_state=`listening`** triggers (sparkles effect, speaker_1 = giver/listener). Recap captures "one expressing worry through sharp words, the other sharing tender vision." Per-utterance ground-truth labels not available post-update; distribution numbers derived from pipeline fused output. |
|
|
| ## Source |
| - Dataset: MELD (Multimodal EmotionLines Dataset) |
| - Source: Friends TV series |
| - Paper: Poria et al., ACL 2019 |
| - Each WAV is a full dialogue concatenated from per-utterance MP4 clips |
| - Audio: 16kHz mono PCM (matches pipeline input format) |
|
|
| ## Usage |
| ```bash |
| # Run pipeline on a single test set |
| python scripts/run_pipeline.py data/meld_test/01_angry_fight.wav |
| |
| # Evaluate all test sets |
| python scripts/evaluate_meld_test.py |
| ``` |
|
|
