Datasets:

liva-ai
/

real-sales

conversation_id int64	role string	speaker_id string	age_range string	gender string	ethnicity string	grew_up_in string
1	Partner	conv01_partner	25-34	Male	White	Mid Atlantic
1	Seller	conv01_seller	25-34	Male	South Asian	West
2	Partner	conv02_partner	25-34	Female	East Asian	Mid Atlantic
2	Seller	conv02_seller	25-34	Female	White	Mountain West
3	Partner	conv03_partner	25-34	Male	White	Non US
3	Seller	conv03_seller	55-64	Female	White	Midwest
6	Partner	conv06_partner	25-34	Male	White	Non US
6	Seller	conv06_seller	45-54	Male	White	Non US
7	Partner	conv07_partner	35-44	Male	White	Non US
7	Seller	conv07_seller	25-34	Male	South Asian	Non US
8	Partner	conv08_partner	25-34	Male	Hispanic	Southern
8	Seller	conv08_seller	25-34	Male	White	Mountain West
9	Partner	conv09_partner	25-34	Female	East Asian	Non US
9	Seller	conv09_seller	25-34	Male	White	Non US

Real Sales Conversations

Real Sales Conversations is a speech dataset of unscripted sales calls between a seller and a prospective partner. Each conversation is captured as two separate single-speaker files, one per participant. Every row in this dataset is one person's isolated channel with their own demographic metadata attached.

Unique aspects of this data

Each participant was recorded on their own device.
Interruptions, backchannels, natural turn-taking behavior present
Real sales conversations, not read or role-played prompts. These are unscripted calls with the disfluencies, hesitation, and turn-taking dynamics of actual business conversation.
Accent and demographic diversity. Speakers span 6 regions of origin including non-US, across multiple age ranges and ethnicities (4 female, 10 male).

What this data is best for

Speech-to-speech and conversational agents in a sales or business context, where realistic turn-taking and interruption handling matter.
ASR on conversational telephony speech. The recordings sit in the acoustic range production call systems actually encounter.

Speaker metadata

Each file carries metadata for the speaker on that channel.

Role (Seller or Partner)
Age range
Gender
Ethnicity (blank where not recorded)
Where they grew up (accent info)

Technical specs

Metric	Value
Sample rate	44,100 Hz / 48,000 Hz (mixed)
Bit depth	16
File format	WAV, PCM, mono, one file per speaker
Mean SNR	53.76 dB
Median RMS	0.031225
Average speech ratio	0.545
Spectral centroid	1.105 kHz
Frequency content	9.81 kHz

Acoustic caveats

These are real-world call recordings, not studio captures

Narrowband source. Despite the 44.1/48 kHz containers, effective bandwidth is ~9.8 kHz with a ~1.7 kHz 95% energy rolloff. This is call-codec audio, upsampled. Only ~1.6% of energy sits above 3.4 kHz.
VoIP noise gating. Non-speech regions are frequently gated to exact digital silence, up to 49% of a file. Reported SNR excludes those gated frames, since including them yields a meaningless ~99 dB.
Mixed sample rates across, and in one case within, conversations.

Downloads last month: 52

Number of rows:

Total file size:

964 MB