Datasets:

liva-ai
/

yapdo-mini

language:
  - en
  - hi
  - ar
  - sw
  - te
  - tl
  - pcm
  - es
license: cc-by-4.0
gated: auto
extra_gated_heading: Access Yapdo-Mini
extra_gated_description: >-
  Please share your contact information to access this dataset. Access is
  granted immediately.
extra_gated_button_content: Agree and access dataset
extra_gated_fields:
  Affiliation:
    type: text
    required: true
  Use case:
    type: select
    options:
      - Research
      - Commercial
      - Education
      - label: Other
        value: other
task_categories:
  - automatic-speech-recognition
  - audio-classification
tags:
  - conversational-speech
  - multilingual
  - accent
  - code-switching
  - multi-speaker
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Yapdo-Mini

Yapdo-Mini is a sample of the Yapdo dataset, a conversational speech corpus drawn from 109,804 hours of recordings from 17,008 speakers across 67 languages. The source audio is natively recorded with separate speaker channels; the samples here are presented as combined conversations.

Yapdo Data Highlights


Total audio	109,804 hours
Unique speakers	17,008
Languages	67
Format	48 kHz, 16-bit PCM WAV per speaker
Channel separation	Each speaker on a dedicated, time-aligned track
Speech type	Spontaneous, unscripted, multi-party conversations
Code-switching	Yoruba-English, Hindi-English, Swahili-English ("Sheng"), Tagalog-Cebuano, and more
Mean SNR	~33 dB
Median RMS	-26 dBFS

Top 10 Languages (estimated hours)

Language	Hours	Language	Hours
English	31,660	Tagalog	2,014
Hindi	8,412	Spanish	1,651
Arabic	2,427	Nigerian Pidgin	1,382
Swahili	2,075	Tamil	1,288
Hausa	2,074	Cebuano	848

Note: These are estimated hours based on automated language detection. We are in the process of obtaining human-verified language and accent labels. The total number of languages and hours per language/accent are subject to change.

Combined vs. Separated Audio

Each sample in this mini dataset is a combined mix of all speakers. The parent Yapdo corpus stores each speaker on a separate, time-aligned track. Here's what that difference sounds like — a Telugu conversation with 2 speakers:

Combined (all speakers mixed)

Speaker 1 (isolated track)

Speaker 2 (isolated track)

All 12 Samples

#	Language	Speakers	Variety	Transcript
1	sw	2	Nairobi urban	Yes
2	hi	2
3	tl	2	Central Visayas
4	sw	2	Nairobi urban	Yes
5	ar	3	Cairene
6	te	2	Karnataka/Bangalore	Yes
7	es	3	Venezuelan
8	pcm	2	Nigerian English	Yes
9	en	3	Egyptian Arabic	Yes
10	pcm	2	Nigerian English	Yes
11	tl	3	Mindoreño
12	en	3	Indian

Schema

Column	Type	Description
`audio`	`Audio(16kHz)`	Combined multi-speaker audio, 16 kHz mono
`text`	`string`	Timestamped transcript with speaker IDs (human-reviewed where available, otherwise empty). Full human-validated transcripts are available upon request.
`language`	`string`	Primary ISO 639-1 language code
`accent`	`string`	Accent or dialect label (e.g. "Nairobi urban", "Cairene", "Mindoreño")
`relationship`	`string`	Speaker relationship (friends, acquaintances, colleagues, etc.)
`topics`	`string`	Topics discussed
`speech_characteristics`	`string`	Notable audio features (code-switching, laughter, etc.)
`num_speakers`	`int`	Number of speakers in the clip
`duration_s`	`float`	Clip duration in seconds
`rms_dbfs`	`float`	RMS loudness in dBFS
`peak_amplitude`	`float`	Peak sample amplitude (0.0–1.0)
`speech_ratio`	`float`	Fraction of frames containing speech

Usage

from datasets import load_dataset

ds = load_dataset("liva-ai/yapdo-mini", split="train")

for example in ds:
    print(f"{example['language']:>3s} | {example['num_speakers']} speakers | {example['variety']}")
    print(f"   Transcript: {example['text'][:100]}...")
    print()