Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 461871917 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Decider Audio – Speaker Count

Part of the Decider Audio collection.

How many different people speak in a conversation? Each row is a synthetic conversation of 6–9 turns by 1–5 speakers. The answer is one of five options ("1" to "5"). The label is exact by construction, because every speaker is a different cloned reference voice. The content of the turns carries no information about the count, so the task can only be solved from the voices.

10,000 conversations, 112.7 hours, 75,973 turns, 1,592 reference voices.

Split Conversations Hours Voices Turns Rows failing checks Speakers 1 / 2 / 3 / 4 / 5
train 7,000 78.84 998 53,221 0 1,400 / 1,400 / 1,400 / 1,400 / 1,400
validation 1,000 11.31 250 7,572 0 200 / 200 / 200 / 200 / 200
test 2,000 22.53 344 15,180 0 400 / 400 / 400 / 400 / 400

No speaker appears in more than one split. A reference speaker belongs to one split only (the same 80/10/10 holdout as every other Decider Audio dataset). A voice is also left out when it sounds too similar to a voice of another split (ECAPA similarity above 0.7). No turn text appears in more than one split either.

How it is made

  • Speakers. There are 1–5 speakers, balanced within each split. Each conversation uses the least-used voices of its split, and no two of its voices have a reference similarity of 0.6 or higher.
  • Turns. A conversation has 6–9 turns, at least n + 2. Every speaker talks at least once, and nobody talks twice in a row.
  • Text. Each turn is a sentence from the participant opinions of DeepMind's Habermas Machine data (CC BY 4.0), 6–22 words long. Train turns come from its train file. Validation and test turns come from its test file, split by policy question.
  • TTS. One model per conversation, balanced within every speaker count, so the model is no cue.
TTS model Checkpoint Conversations Turns Re-rendered turns
pocket kyutai/pocket-tts 2,000 15,115 141
sopro samuel-vitorino/sopro-v2-turbo 2,000 15,254 20
qwen Qwen/Qwen3-TTS-12Hz-0.6B-Base 2,000 15,221 81
omni k2-fsa/OmniVoice 2,000 15,178 67
chatterbox ResembleAI/chatterbox 2,000 15,205 41
  • Mixing. Each turn is trimmed (−45 dB) and set to −23 dB RMS, so loudness is no cue. Turns are separated by 0.35–0.9 s of silence and never overlap. The audio is 24 kHz mono FLAC, 18.9–115.5 s long (mean 40.6 s).

Checks

Every rendered turn is checked twice.

  • Text. The ASR (Systran/faster-whisper-base.en) must match the text. The WER must be at most 0.5, the speech at most 30 characters per second, and fewer than 3 extra words may appear.
  • Speaker. The check uses ECAPA embeddings (speechbrain/spkrec-ecapa-voxceleb).
    • The turn's similarity to the centroid of its speaker's other turns must be at least 0.4.
    • That similarity must exceed the similarity to every other speaker's centroid by at least 0.1.
    • The similarity to any other speaker must stay below 0.55.

A failing turn is rendered again with the same voice and text and a new seed, up to 5 renders. The best render is kept. A second repair pass rendered 118 turns again with a new seed, and with a new voice for their speaker where that still failed. 3 of them passed the word check but leave out part of their text. A turn 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 turn that leaves out part of its text does not pass this way.

  • 350 turns use a re-render.
  • 0 turns in 0 conversations still fail a check. They are kept and marked passes_checks = false, at row and turn level.
  • Mean turn WER is 0.023. 78.2% of the turns are transcribed exactly.
  • Speaker similarity medians are 0.83 to the own speaker and 0.17 to the closest other speaker. The 1st percentile of the own similarity is 0.58 and the 99th percentile of the other similarity is 0.46.

Embedding baseline. Average-linkage clustering of the turn embeddings at cosine distance 0.6 finds the right count in 97.2% of the conversations. By speaker count 1 / 2 / 3 / 4 / 5 it is right in 99.9% / 99.5% / 98.6% / 95.5% / 92.7%. The result is stored as ecapa_cluster_count. This is a diagnostic and not a ceiling, and it used the true turn boundaries.

Fields

Field Meaning
id speaker_count-<split>-<index>
audio The conversation (24 kHz mono FLAC)
instruction, choices, choice_keys The question, the options "1"–"5" and their keys A–E
answer, answer_index, n_speakers The correct key, its index and the speaker count
n_turns, turns Turn count, and a JSON list per turn with speaker index, start and end in seconds, text, voice id, TTS seed, render attempt, ASR text and WER, the speaker similarities, and passes_checks
speaker_voice_ids Reference voice id per speaker
tts_model The TTS model of the conversation
duration_seconds, render_attempt, passes_checks Length, highest render attempt used, all checks passed
ecapa_cluster_count Speakers found by the clustering baseline
source_dataset, source_ids Habermas Machine and the source record per turn

License and attribution

  • The turn 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), Chatterbox (MIT) 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 turns field contains speaker indices only. There are no names.

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

Downloads last month
29

Collection including SebastianBodza/decider-audio-speaker-count