charcot-ai/marimba-bench

Marimba Bench

Curated public audio used to train and evaluate the readouts of Marimba, a general-purpose model that answers typed questions about audio (turn end, voice activity, speaker, language, and other decisions). Every row is derived from a public dataset or synthesized, and keeps the license of its source.

• 25 configs, 90,112 rows, 136.38 hours.
• Mono audio, 16 kHz, stored in Parquet shards.
• Release tag: v1.2.0.

Configs and splits

Splits follow the official upstream splits. Common Voice uses a deterministic per-language sample of train; Speech Commands uses a deterministic hash sample within each official split.

Config	Source	Train	Validation	Test	Hours
ami_segments	AMI	7,314	982	835	7.27
ami_vad	AMI	4,352	576	512	12.09
common_voice_en	Common Voice	3,750	—	—	5.78
common_voice_es	Common Voice	3,852	—	—	5.75
common_voice_fr	Common Voice	3,788	—	—	5.62
common_voice_it	Common Voice	3,750	—	—	6.41
common_voice_pt	Common Voice	3,750	—	—	4.90
common_voice_ro	Common Voice	3,750	—	—	4.21
fleurs_ar_eg	FLEURS	1,161	144	283	3.36
fleurs_cmn_hans_cn	FLEURS	1,187	150	349	3.58
fleurs_de_de	FLEURS	1,184	147	347	3.61
fleurs_en_us	FLEURS	1,225	150	350	3.63
fleurs_es_419	FLEURS	1,187	150	348	3.66
fleurs_fr_fr	FLEURS	1,198	137	332	3.63
fleurs_hi_in	FLEURS	1,198	126	265	3.41
fleurs_it_it	FLEURS	1,194	147	346	3.64
fleurs_ja_jp	FLEURS	1,170	135	321	3.56
fleurs_ko_kr	FLEURS	1,190	129	270	3.49
fleurs_pt_br	FLEURS	1,175	147	349	3.67
fleurs_ru_ru	FLEURS	1,235	150	344	3.73
kokoro_pii	Synthetic (Kokoro-82M)	72	—	—	0.12
musan	MUSAN	5,726	—	—	12.72
speech_commands	Speech Commands v0.02	4,350	4,350	4,350	3.56
voxconverse_segments	VoxConverse	—	1,425	3,108	3.65
voxconverse_vad	VoxConverse	—	4,800	4,800	21.33
Total		58,758	13,845	17,509	136.38

Each FLEURS config (fleurs_ar_eg, fleurs_cmn_hans_cn, fleurs_de_de, fleurs_en_us, fleurs_es_419, fleurs_fr_fr, fleurs_hi_in, fleurs_it_it, fleurs_ja_jp, fleurs_ko_kr, fleurs_pt_br, fleurs_ru_ru) has train, validation, and test splits. Two FLEURS Arabic rows (one in train, one in test) were skipped because their audio failed to decode.

Contents

• *_vad: 8.0 s windows from AMI and VoxConverse with 80 frame labels (100 ms each) in frame_labels.
• *_segments: 2–4 s single-speaker segments from AMI and VoxConverse, with speaker_id, session_id, start, and end.
• common_voice_* and fleurs_*: read speech of at most 8.0 s, with language and text.
• speech_commands: clips of at most 1.0 s with the keyword label and speaker_id.
• musan: 8.0 s windows, up to 4 per file (music 2,634; noise 1,391; speech 1,701).
• kokoro_pii: 72 synthetic utterances generated with Kokoro-82M from templates containing fabricated identifiers (for example, phone numbers) in identifier. No real personal data.

All configs share one schema: audio (16 kHz); string fields id, split, sentence, text, speaker_id, language, config, label, file_id, session_id, voice, template_id, identifier, identifier_type, path; float fields start, end, duration_s; window_index (int32); frame_labels (80 × int8). Fields not used by a config are empty.

Sources and licenses

Source	Revision	License
AMI Meeting Corpus (diarizers-community/ami)	8cdaae2e	CC BY 4.0
VoxConverse (diarizers-community/voxconverse)	3acfa1b4	CC BY 4.0
FLEURS (google/fleurs)	168de341	CC BY 4.0
Speech Commands (google/speech_commands)	v0.02	CC BY 4.0
Mozilla Common Voice, resampled to 16 kHz	03c7beb9	CC0 1.0
MUSAN (OpenSLR 17)	SHA-256 86d1061c…	CC BY 4.0
Kokoro-82M (hexgrad/Kokoro-82M), synthetic speech	f3ff3571	Apache 2.0 (model)

The Marimba turn-end readout is trained and tested on Smart Turn v3.2 (CC BY 4.0), which is not included here; use it from its source.

Redistribution

The audio is derived from third-party datasets and is not relicensed. Each row remains under its upstream license and terms; attribution to the original authors is required for CC BY 4.0 sources. Do not redistribute this audio under a different license; obtain it from the original sources when your use requires it. No private or production recordings are included.

Attribution

The following sources are distributed under Creative Commons Attribution 4.0 International (CC BY 4.0):

• AMI Meeting Corpus: Carletta et al. (2005), The AMI Meeting Corpus: A Pre-announcement.
• VoxConverse: Chung et al. (2020), Spot the Conversation: Speaker Diarisation in the Wild.
• FLEURS: Conneau et al. (2022), FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech.
• Speech Commands: Warden (2018), Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition.
• MUSAN: Snyder, Chen, and Povey (2015), MUSAN: A Music, Speech, and Noise Corpus. Individual music and noise files retain their upstream licenses; review the original corpus license files before redistribution.

Changes: resampled to 16 kHz mono; windowed/segmented as described above.

Mozilla Common Voice is provided under CC0 1.0; do not attempt to identify speakers.

Citation

@misc{marimba2026,
  title  = {marimba: A System One Model for Speech},
  author = {Solano Amadori},
  year   = {2026}
}