SebastianBodza's picture
Initial release
2baeabc
|
Raw History Blame Contribute Delete
11 kB
metadata
license: other
license_name: mixed-see-card
language:
  - en
task_categories:
  - audio-classification
  - question-answering
tags:
  - audio
  - speech
  - speaker-verification
  - speaker-diarization
  - audio-quality
  - reverberation
  - synthetic
  - decider-audio
pretty_name: Decider Audio – audio-native decision tasks
configs:
  - config_name: speaker_attribution
    data_files:
      - split: train
        path: data/speaker_attribution/train-*
      - split: validation
        path: data/speaker_attribution/validation-*
      - split: test
        path: data/speaker_attribution/test-*
  - config_name: same_speaker
    data_files:
      - split: train
        path: data/same_speaker/train-*
      - split: validation
        path: data/same_speaker/validation-*
      - split: test
        path: data/same_speaker/test-*
  - config_name: reverb
    data_files:
      - split: train
        path: data/reverb/train-*
      - split: validation
        path: data/reverb/validation-*
      - split: test
        path: data/reverb/test-*
  - config_name: quality
    data_files:
      - split: train
        path: data/quality/train-*
      - split: validation
        path: data/quality/validation-*
      - split: test
        path: data/quality/test-*
  - config_name: quality_pair
    data_files:
      - split: train
        path: data/quality_pair/train-*
      - split: validation
        path: data/quality_pair/validation-*
      - split: test
        path: data/quality_pair/test-*

Decider Audio – audio-native decision tasks

Part of the Decider Audio collection.

Five multiple-choice tasks whose answer is in the sound, not in the words. The model hears one or two recordings and must pick an option. None of the tasks can be solved from a transcript. Labels are exact by construction, because the voices, the simulated rooms and the applied degradations are known. Each row is rendered on its own, and no clip appears in two rows.

21,500 rows, 70.9 hours of audio, 51,604 rendered clips, 1,592 reference voices.

Task (config) Question
speaker_attribution 2–4 speakers each introduce themselves by name ("Hi, this is Alex. …"), then talk in random order. The question is either Who says ""? or Who speaks last?.
same_speaker Two recordings separated by a beep. Are both spoken by the same person? Half of the "different" pairs are hard negatives, made from the 5 most similar voices of the split.
reverb Two recordings by one voice. In which one does the room sound more reverberant?
quality Which degradation does this recording have? Choices are clean, noise, hum, clipping, telephone band, MP3 at 8–16 kbit/s, and dropouts.
quality_pair The same clip at a mild and a strong level of one degradation. Which recording has the better quality?
Task Train Validation Test Hours Choices Rows failing checks
speaker_attribution 3,000 300 1,000 34.0 2–4 0
same_speaker 3,000 300 1,000 10.0 2 0
reverb 3,000 300 1,000 12.6 2 1
quality 3,000 300 1,000 4.2 7 0
quality_pair 3,000 300 1,000 10.1 2 0
from datasets import load_dataset

ds = load_dataset("SebastianBodza/decider-audio-native", "speaker_attribution", split="test")

How it is made

  • Texts. The text source is single sentences (6–22 words) from participant opinions in the Habermas Machine data (CC BY 4.0). Train uses train opinions. Validation and test use test opinions, split by policy question. Sentences with a gendered self-reference ("as a mother") are left out.

  • Voices. The voices come from a private collection of English reference recordings, the same pool as decider-audio-speaker-count.

    • No speaker appears in more than one split. A voice is also left out when it sounds too similar to a voice of another split (ECAPA similarity above 0.70).
    • Two voices in one row, and the two voices of a "different" pair, stay below 0.6.
  • No text shortcut.

    • Attribution sentences are assigned to speakers at random, so a speaker's sentences share no topic or stance.
    • Content turns follow in random order, with repeats allowed, so the speaking order does not follow from the introduction order.
    • The choices are shuffled independently of the introduction order.
    • The answer is uniform over introduction positions and over answer keys (see the balance tables below).
  • One model per row. The four TTS models are balanced within every task, split and label.

  • Levels. Every turn and recording is set to -23.0 dB RMS, so loudness is no cue.

  • Rooms. The reverb task uses rooms simulated with pyroomacoustics (image source method, inverse Sabine absorption).

    • The target RT60 is 0.25–1.8 s, and room sizes grow with RT60.
    • The two rooms of a row differ by at least 1.6×.
    • The speaker-to-microphone distance is 1.0–1.6 m, the same in both rooms.
    • Both recordings use the same voice, so the room baked into the reference recording cancels out.
    • The measured RT60 order agrees with the answer in 4,300 of 4,300 rows.
  • Degradations. The quality and quality_pair tasks use these degradations.

    • Noise at SNR 0–15 dB.
    • Hum at 50 Hz with harmonics, SNR 0–12 dB.
    • Clipping with 4–10× overdrive.
    • Telephone with a 300–3400 Hz band and 8 kHz resampling.
    • MP3 at 8 or 16 kbit/s.
    • Dropouts in 6–20 % of 40 ms chunks.

    In quality_pair, the mild and the strong level are far apart (for example SNR 25–35 vs. 0–8 dB, MP3 64–96 vs. 8–16 kbit/s).

TTS model Checkpoint Rows Clips Re-rendered clips
pocket mlx-community/pocket-tts 5,387 12,947 318
sopro samuel-vitorino/sopro-v2-turbo 5,381 12,889 213
qwen Qwen/Qwen3-TTS-12Hz-0.6B-Base 5,373 12,888 219
omni k2-fsa/OmniVoice 5,359 12,880 67

Checks

Every clip was transcribed with Systran/faster-whisper-base.en. A clip fails if any of these holds.

  • The WER is above 0.5.
  • It is spoken faster than 30 characters per second.
  • It adds words.
  • It is an introduction and the name is not recognized.

Speaker tasks also check ECAPA embeddings (speechbrain/spkrec-ecapa-voxceleb). A clip fails if any of these holds.

  • Its similarity to the same speaker's other clips is below 0.4.
  • That similarity is not at least 0.1 above its similarity to another speaker.
  • Its similarity to another speaker reaches 0.55.

A failing clip is rendered again with the same voice and text and a new seed, up to 5 renders. A second repair pass rendered 105 clips again with a new seed, and with a new voice for their speaker where that still failed. 5 of them passed the word check but leave out part of their text. A clip now also passes when its only problem is the word error rate or extra words and its transcript sounds like the text. This check compares phonemes after numbers and units are spelled out, so "topup" and "top up" or "1960" and "nineteen sixty" match. A clip that leaves out part of its text does not pass this way.

  • Re-rendered clips: 817.
  • Clips still failing: 1. Their rows have passes_checks = false.
  • Mean WER: 0.038.
  • Median render similarity in same_speaker, by pair type: easy 0.10 / hard 0.28 / same 0.68.

Answer balance

Task Split Answer keys (count)
speaker_attribution train A 1,100, B 1,069, C 585, D 246
speaker_attribution validation A 93, B 114, C 74, D 19
speaker_attribution test A 379, B 346, C 188, D 87
same_speaker train A 1,500, B 1,500
same_speaker validation A 150, B 150
same_speaker test A 500, B 500
reverb train A 1,500, B 1,500
reverb validation A 150, B 150
reverb test A 500, B 500
quality train A 429, B 429, C 429, D 429, E 428, F 428, G 428
quality validation A 43, B 43, C 43, D 43, E 43, F 43, G 42
quality test A 143, B 143, C 143, D 143, E 143, F 143, G 142
quality_pair train A 1,500, B 1,500
quality_pair validation A 150, B 150
quality_pair test A 500, B 500

Attribution, introduction position of the correct speaker (uniform means the order carries no information).

Speakers Answer is the 1st / 2nd / 3rd / 4th introduced speaker
2 699 / 735
3 482 / 486 / 465
4 323 / 349 / 394 / 367

Known limits.

  • All speech is synthetic (voice clones). Clones can sound less distinct than the real speakers.
  • The rooms are simulated shoeboxes, and the degradations are synthetic.
  • MP3 at 8 kbit/s also narrows the bandwidth, so codec and telephone can sound alike. They differ in their artifacts and their low end.
  • The reference recordings carry their own room and microphone. reverb and quality_pair compare two recordings of one voice, so this cancels out. In quality, a "clean" clip can still carry the reference's own character.

Fields

Field Meaning
audio One FLAC (24 kHz, mono). Pair tasks contain both recordings, separated by 0.6 s silence, a 0.2 s beep and 0.6 s silence.
instruction, choices, choice_keys The question and its options (A, B, …).
answer, answer_index The correct key and its index.
tts_model, tts_runtime The TTS model of every clip in the row, and the runtime it used.
voice_ids Reference voice ids, one per speaker.
passes_checks Every clip passed the text and speaker checks (and, for reverb, the measured RT60 order agrees with the answer).
details JSON with the task facts and per-clip checks. Attribution has names, speaker of each turn and the question. Same speaker has pair type and the similarity of the references and of the renders. Reverb has target and measured RT60, room size and distance. Quality has degradation and severity. Spans give start and end of each turn or recording in seconds.

License and attribution

  • The texts derive from the Habermas Machine data by Google DeepMind (CC BY 4.0).
  • The audio is synthetic. It was rendered with Pocket TTS (CC BY 4.0, attribution), Sopro (Apache-2.0), Qwen3-TTS (Apache-2.0) and OmniVoice (reported as CC BY-NC 4.0, verify on k2-fsa/OmniVoice).
  • Reference voices come from a private collection and are not redistributed. Voice ids are anonymous hashes.

The dataset is marked license: other. Do not use it commercially.