---
license: other
license_name: mixed-cc-by-4.0-and-higgs-tts-3
task_categories:
  - audio-to-audio
language:
  - en
tags:
  - speech
  - full-duplex
  - emotion
  - conversational
size_categories:
  - 10K<n<100K
pretty_name: Emotionplex duplex multispeaker
dataset_info:
  - config_name: full
    features:
      - name: path
        dtype: audio
      - name: duration
        dtype: float64
    splits:
      - name: train
        num_examples: 22800
      - name: validation
        num_examples: 1200
  - config_name: preview
    features:
      - name: audio
        dtype: audio
      - name: duration
        dtype: float64
      - name: text_prompt
        dtype: string
      - name: id
        dtype: string
    splits:
      - name: train
        num_examples: 8
configs:
  - config_name: preview
    default: true
    data_files:
      - split: train
        path: preview.jsonl
  - config_name: full
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: val.jsonl
---

# Emotionplex duplex multispeaker

About 55 hours of English full-duplex conversations for finetuning [PersonaPlex](https://huggingface.co/nvidia/personaplex-7b-v1). Each clip is one user turn and one agent reply. Speech was synthesized with [Boson AI Higgs Audio v3 TTS 4B](https://huggingface.co/bosonai/higgs-tts-3-4b) (Higgs TTS 3). The agent was trained to follow:

`Answer with a {emotion} tone.`

where `{emotion}` is `happy`, `sad`, `angry`, or `neutral`.

Speech time by emotion, from each turn's start and end. Padding and the gap between speakers are not included. The channels overlap, so this sum is longer than the stereo timeline. Each speaker has 6,000 turns in every emotion, 1,500 clips for each of the 16 user-to-agent pairs.

| | Happy | Sad | Angry | Neutral |
|---|---:|---:|---:|---:|
| User | 7h53m | 7h28m | 7h02m | 5h59m |
| Agent | 7h28m | 8h01m | 6h18m | 5h14m |

Total speech time is 55h23m.

This dataset is not affiliated with NVIDIA or Boson AI. See `NOTICE`.

The dataset page opens on the `preview` subset: eight full stereo clips, one per emotion pair. The rest are in the `full` subset. Those show “Not supported with pagination yet” in the viewer.

## Layout

Stereo WAV at 24 kHz. Left channel is the agent (`SPEAKER_BROKER`). Right channel is the user (`SPEAKER_CLIENT`). A `.json` file with the same stem sits next to each WAV.

| Path | What it is |
|---|---|
| `stereo_wav/<shard>/` | Stereo clips and sidecars (`text_prompt`, alignments, voice, emotion). Shards stay under the Hub limit of 10,000 files per directory. |
| `turns/<shard>/` | The user and agent mono takes packed into each stereo clip |
| `train.jsonl`, `val.jsonl` | `{"path": "stereo_wav/000/<id>.wav", "duration": ...}` |
| `dialogues.jsonl`, `users.jsonl` | Plain text of the lines before TTS |
| `voice_prompts/` | LibriTTS-R reference clips used as the PersonaPlex voice prompt |
| `roomtone/classroom.wav` | Room tone mixed under the user channel |

Paths in `train.jsonl`, `val.jsonl`, sidecar `wav`, and `turns[].path` are relative to this dataset root. `voice_prompt` and `voices.*.path` are relative to the stereo file, for example `../../voice_prompts/103.wav`, because the PersonaPlex trainer joins a relative voice path onto the WAV's directory (`stereo_wav/<shard>/`).

`sphn` opens `train.jsonl` paths from the process working directory. Train with this dataset directory as the working directory, or expand those paths to absolute paths on your machine.

## License

Mixed. Synthetic speech was produced with Higgs TTS 3. Voice clips are LibriTTS-R (CC BY 4.0). Room tone is Freesound 212202 by klankbeeld (CC BY). Details and required credits are in `NOTICE`.