liva-ai/social-streaming-samples

Description of Data Source

A social entertainment app where people stream content together and talk freely over live voice calls. Friends and other participants can join shared sessions to socialize and have spontaneous conversations. Audio is recorded separately on each device, producing time-aligned, channel-separated tracks.

Technical Analysis

Measure	Value
Sample rate	48 kHz
Bit depth	16-bit PCM
File format	WAV (stereo, one person per channel)
Processing	Lossy: recorded by the app as Opus at about 15 kbps, with noise suppression and echo cancellation
Bandwidth	Wideband: speech content up to about 8 kHz, some up to 12 kHz

These metrics vary from conversation to conversation. Contact us for a breakdown at the quality level you need. Pricing depends on that level.

Measure	Range
Speech level (RMS)	about -29 to -15 dBFS
Speech ratio	median 0.38 per conversation
Audio bleed	almost none (under 1 % of windows)
Spectral centroid	about 0.6-1.3 kHz

Audio processing. The app records audio using lossy Opus compression at approximately 15 kbps and provides decoded 48 kHz / 16-bit WAV files. The recordings carry speech content up to about 8 kHz, with some up to 12 kHz. Noise suppression and echo cancellation reduce background noise and audio bleed.

Languages

Distribution of conversation hours by language.

Language	Share of hours
Arabic	31.0 %
Russian	15.0 %
Spanish	13.8 %
Hindi / Urdu	13.7 %
Portuguese	9.9 %
Turkish	7.6 %
English	2.6 %
Azerbaijani	1.7 %
German	0.7 %
Kurdish	0.6 %
Indonesian	0.5 %
Unknown	2.9 %

Distribution of conversation hours by country within each language. Speaker locations are estimated using IP addresses or profile locations for VPN users.

Language	Countries
Arabic	Saudi Arabia 17 %, Kuwait 15 %, Egypt 14 %, Iraq 9 %, Algeria 7 %, Jordan 7 %, Morocco 5 %
Russian	Russia 77 %, Belarus 4 %, Ukraine 3 %, Uzbekistan 3 %, Kazakhstan 2 %
Spanish	Mexico 22 %, Peru 15 %, Argentina 13 %, Colombia 10 %, Chile 8 %, Ecuador 5 %, United States 5 %
Hindi / Urdu	India 53 %, Pakistan 35 %
Portuguese	Brazil 97 %
Turkish	Turkey 90 %, Azerbaijan 3 %

Speakers per Conversation

Distribution of conversation hours by number of speakers.

Speakers	Share of hours
2	74.7 %
3	6.2 %
4-5	5.7 %
6-10	7.0 %
11+	6.4 %

Conversation Categories

Distribution of two-person conversation hours by category and language. Categories were defined and conversations classified from transcripts using an LLM.

	English	Hindi	Portuguese	Russian	Spanish	Turkish	All
Gaming and media	31 %	10 %	46 %	44 %	36 %	33 %	33 %
Romance and flirting	27 %	12 %	12 %	7 %	17 %	21 %	18 %
Casual storytelling	10 %	6 %	7 %	11 %	13 %	9 %	9 %
Humour and banter	7 %	12 %	5 %	10 %	4 %	7 %	8 %
Emotional support	7 %	12 %	4 %	6 %	5 %	3 %	6 %
Disagreement	2 %	11 %	5 %	3 %	6 %	3 %	5 %
Other	6 %	4 %	5 %	3 %	3 %	6 %	5 %

Relationships

Distribution of conversation hours by relationship type and language. Relationships were classified from transcripts using an LLM.

	English	Hindi	Portuguese	Russian	Spanish	Turkish	All
Close friends	53 %	75 %	73 %	88 %	66 %	66 %	67 %
Romantic partners	31 %	5 %	23 %	6 %	27 %	30 %	22 %
Acquaintances	11 %	19 %	4 %	7 %	6 %	4 %	9 %

Accents

Distribution of English conversation hours by estimated accent. At least one participant in each conversation has the specified accent. Accents are inferred from IP addresses or profile locations for VPN users.

Accent	Share of hours
South Asian English	21.7 %
General American	17.7 %
Latin American influenced English	12.2 %
Southern American English	10.1 %
Middle Eastern influenced English	7.8 %
Maritime Southeast Asian English	6.8 %
Western European influenced English	2.8 %
Canadian English	2.6 %
Northern British English	2.2 %
Eastern European influenced English	2.2 %
Southern British English	2.2 %
Mainland Southeast Asian influenced English	1.6 %
Australian English	1.2 %
Southern European influenced English	1.2 %
North African influenced English	1.0 %
Other (each under 1 %)	6.7 %