Expressivity / Emotional Intelligence

Vocal Bursts & Disfluencies

Non-verbal sounds stay in the transcript instead of being stripped out. Vocal bursts like [laughing], [sigh] and [gasp], fillers like [um] and [hmm], and delivery spans like <whisper> and <elongated> are tagged at the exact position they occur, so models learn when they happen, not just how they sound.

11Kbursts

69Kfillers

Request samples

Specification

Bursts: [laughing] [sigh] [gasp] [cough] [crying] [groan]
Fillers: [um] [uh] [hmm] [mm] [ah] [ohh]
Style spans: <whisper> <elongated>
Placement: Inline, at the exact word position

Delivery layout

ampllab-speech-to-speech-batch-<timestamp>/
session_<id>/
scene_<n>/
take-<n>/
<speaker-a>.flacCH1, isolated
<speaker-b>.flacCH2, isolated
merged-<take>.flacmixed
annotation-<id>.jsontranscript + labels

Vocal bursts

[laughing]
[sigh]
[gasp]
[cough]
[crying]
[groan]
[singing]

Fillers

[um]
[uh]
[hmm]
[mm]
[ah]
[ohh]
[haan]

Style spans

<whisper>...</whisper>
<elongated>...</elongated>

take-<n>/annotation-<id>.jsonIllustrative values

1{2  "script": {3    "category": "Banking",4    "sub_category": "Card Servicing"5  },6  "scene": {7    "id": "scene_2",8    "name": "Declined card at checkout"9  },10  "take": {11    "uid": "7f3c9e21-...",12    "seq": 4,13    "duration": 287.4214  },15  "speakers": {16    "a41d07c2-...": {17      "role": "Priya (Card Services Agent)"18    },19    "e9b2f6a8-...": {20      "role": "Daniel (Customer)"21    }22  },23  "transcription": [24    {25      "speaker": "e9b2f6a8-...",26      "start_time": 3.05,27      "end_time": 7.48,28      "transcript": "[sigh] Hi, my card got declined twice <overlap>this morning</overlap>.",29      "emotion": "Frustrated",30      "languages": "English"31    },32    {33      "speaker": "a41d07c2-...",34      "start_time": 7.1,35      "end_time": 11.86,36      "transcript": "<overlap>Oh no,</overlap> let me take a look [mm]. Can you confirm the last four digits?",37      "emotion": "Calm",38      "languages": "English"39    }40  ]41}

Sample pack

A curated set of takes from this dataset, with audio and annotation files in the delivery format.