liva-ai/yapdo-2

Yapdo Sample Data

This dataset card shows details about the Yapdo conversational speech corpus by Liva AI (YC S25). This dataset card details information for 30,000+ hours of recordings from 8,000+ speakers across 17 languages, with the rest of the hours still undergoing QA. The source audio is natively recorded with separate speaker channels; the samples here are presented as combined conversations.

The strength of this dataset is its naturalness. Recorded among friends "in-the-wild," it preserves the spontaneity of real dialogue and supports the development of better conversational AI. It also reflects friendly interactions across cultures and captures realistic turn-taking dynamics, which are essential for training models that sound natural.

Dataset Configs

Each config maps to a single directory of audio files with an accompanying metadata.jsonl.

Group	Configs	Description
Languages	languages_ar, languages_arz, languages_bn, languages_es, languages_fr, languages_gu, languages_ha, languages_hi, languages_ig, languages_it, languages_kn, languages_pa, languages_pcm, languages_si, languages_sw, languages_ta, languages_te, languages_tl, languages_ur, languages_yo	A subset of our multilingual speech organized by language code. Metadata: nisqa
Accents	accents_Arabic_influenced_English, accents_East_African_English, accents_Filipino_English, accents_General_American, accents_Indian_English, accents_Nigerian_English, accents_South_African_English, accents_West_African_English	Accent-stratified English speech across 8 accent groups. Metadata: nisqa
Full-duplex	full_duplex_backchanneling, full_duplex_interruption_detection, full_duplex_pause_handling, full_duplex_turn_taking	Recordings with full-duplex benchmark characteristics (backchanneling, interruption, pause handling, turn-taking)
Paralinguistics	paralinguistics	Paralinguistic elements (laughing, singing, elongated words, etc.). For the sake of displaying these elements, we identified emotion-rich segments using Empathic-Insight-Voice-Small since we observed that emotion-rich segments were likely to contain such elements. Metadata: emotion, peak_score, avg_mos
Random 50	random_50	50 completely random samples from our QA'd pool. Metadata: nisqa, language, accent, speech_ratio
Gaming	gaming_assistant, gaming_companion	Gaming conversations: helper-type conversations vs. companion. Metadata: game, summary

Dataset Overview

Total audio	30,381 hours
Unique conversations	30,444
Unique speakers	8,026
Languages	~17
Speakers per conversation	2–13 (avg 2.7)
Conversation duration	19s – 24.4 hrs (avg ~58 min)
Code-switching	28.1% of conversations
Speech type	Spontaneous, unscripted, multi-party conversations
Quality score (NISQA)	1.8 - 4.8 (avg 2.5)
Common topics	Video games, daily life (jobs, school, relationships, earning money)

---

Languages

Language labels for each conversation were reviewed by a native human speaker.

Monolingual Conversations

17 languages with over 10 hours of monolingual conversation data. The below includes an estimation on the number of hours.

Language	Code	Conversations	Hours
English	en	13,339	10,860.1
Egyptian Arabic	arz	1,647	1,537.5
Spanish	es	1,016	1,388.1
Swahili	sw	1,358	850.3
Nigerian Pidgin	pcm	744	698.8
Arabic	ar	559	447.4
Hindi	hi	795	434.1
Tagalog	tl	150	229.9
Tamil	ta	138	163.8
Yoruba	yo	181	147.4
Italian	it	255	145.4
Hausa	ha	188	131.7
French	fr	32	49.0
English (alt)	eng	29	32.1
Igbo	ig	33	18.9
Telugu	te	15	12.0
Kannada	kn	14	10.0

Code-Switching Conversations

25 language combinations with over 10 hours of code-switching data, spanning 28.1% of all conversations. The below iincludes an estimation of the groups and hours.

Language Group	Conversations	Hours
English + Nigerian Pidgin	4,206	5,018.7
English + Tagalog	1,537	2,298.9
Cebuano + English + Tagalog	727	1,070.2
English + Swahili	447	504.7
English + Yoruba	238	235.0
English + Hausa	174	232.7
English + Nigerian Pidgin + Yoruba	88	103.9
English + Hindi	148	91.4
Hausa + Swahili	70	92.3
English + Hiligaynon + Tagalog	52	79.1
English + Hausa + Swahili	36	63.7
Nigerian Pidgin + Yoruba	66	60.2
Arabic + English	58	53.3
Hindi + Urdu	41	50.5
English + Tamil	42	44.1
English + Igbo + Nigerian Pidgin	17	33.0
English + Hausa + Nigerian Pidgin	21	32.6
English + Spanish	22	28.3
English + Igbo	29	28.1
English + Telugu	25	25.2
Cebuano + English + Hiligaynon + Tagalog	12	23.0
Igbo + Nigerian Pidgin	19	21.7
Nigerian Pidgin + Swahili	16	19.9
English + Nigerian Pidgin + Swahili	10	15.4
Hausa + Nigerian Pidgin	11	11.1

---

Labels

Language labels were assigned at the speaker-track level by native speakers who reviewed each individual track within a conversation. A single conversation may carry multiple language labels when speakers use different languages. Accent labels are derived from each speaker's self-reported city of origin, providing a natural geographic proxy for dialect and accent variation.

---

Technical Analysis

Sample rate	48 kHz
Bit depth	16-bit PCM
File format	WAV
Mean SNR	~33 dB
Median RMS	-26 dBFS
Average speech ratio	0.35
Spectral centroid	~0.66 kHz
Frequency content	3.3 kHz (averaged over 10–30 second clips)

Combined vs. Separated Audio

Each sample in this 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)

---

Audio Artifacts

Source audio passes through a Discord/Opus VoIP pipeline.

Artifact	Prevalence
Dropouts / packet loss	98.6%
Bandwidth ceiling (< 4 kHz)	97.2%
Clicks / pops	93.6%
Mains hum (50/60 Hz)	82.4%
Silence / dead air	34.6%
Frame repetition	18.2%
Echo	15.2%
Low signal level	5.8%
Onset transients	5.2%
Clipping	0.6%

---