Conversational audio · Quarry Docs

Real calls from contact centres, BPOs and platforms across healthcare, finance and e-commerce / food delivery, de-identified at the source. Each call is one record: an audio file, an optional transcript, and its metadata.

Audio

Item	Spec
Format	WAV or FLAC, lossless
Sample rate	The source's native rate: 8 kHz telephony or higher. Never resampled upward.
Channels	Agent and customer on separate channels where the source recorded them; otherwise mono, flagged in metadata
Redaction	Spans with personal data are replaced with a tone or silence. See audio and transcript redaction.
Naming	One file per call, named by `record_id`

Transcript format

Transcripts are optional per listing. When present, they're time-aligned and speaker-labelled, in the original script plus a romanised copy. Here's the example call on a timeline; hover a segment to read it. Hatched spans were redacted.

agent · 0.0–5.6sNamaste, main aapke health insurance renewal ke baare mein call kar raha hoon.

Each transcript is one JSON file per call, transcripts/<record_id>.json:

{
  "record_id": "qr_01HZX4T9K2M7",
  "language": ["hi", "en"],
  "code_mixed": true,
  "segments": [
    {
      "speaker": "customer",
      "start": 6.1,
      "end": 9.0,
      "text": "हाँ, बोलिए। प्रीमियम कितना होगा?",
      "script": "Deva",
      "text_romanised": "Haan, boliye. Premium kitna hoga?",
      "redactions": []
    },
    {
      "speaker": "customer",
      "start": 17.2,
      "end": 22.5,
      "text": "ठीक है, आप मुझे [PHONE_NUMBER] पर डिटेल्स भेज दीजिए।",
      "script": "Deva",
      "text_romanised": "Theek hai, aap mujhe [PHONE_NUMBER] par details bhej dijiye.",
      "redactions": [{ "type": "PHONE_NUMBER", "start": 19.6, "end": 21.3 }]
    }
  ]
}

Field	Type	Meaning
`segments[].speaker`	enum	`agent`, `customer`, or `ivr` for an automated leg
`segments[].start`, `end`	number	Seconds from the start of the audio
`segments[].text`	string	The utterance in its original script, with redaction tags
`segments[].script`	string	ISO 15924 script code, such as `Deva`, `Taml`, `Latn`
`segments[].text_romanised`	string	The same utterance in Latin script
`segments[].redactions`	array	Each redacted span: entity type, start and end in seconds

Metadata

Field	Type	Values
`language`	string[]	BCP 47 codes, such as `hi`, `ta`, `en-IN`
`code_mixed`	boolean	True when the call switches languages
`domain`	enum	`healthcare`, `finance`, `ecommerce_food_delivery`
`call_type`	string	Such as `health_insurance_sales`, `appointment_scheduling`, `card_collections`, `order_issue`
`duration_s`	number	Length of the audio in seconds
`year`	integer	The year the call was recorded
`channels`	enum	`stereo_split` or `mono`
`sample_rate_hz`	integer	Native sample rate
`outcome`	enum, optional	`sale`, `no_sale`, `resolved`, `escalated`, `promise_to_pay`, where the source records it

Language coverage is listed per dataset on its deal sheet, never site-wide. The full field list is in the metadata dictionary.