Audio dataset, FLAC
Conversational speech
Spontaneous conversation and monologue recorded on contributors' own phones at home, at work, in markets and on calls. Each utterance has one metadata row covering language, setting, speaker, capture settings and consent.
Request access
This is an example dataset. The figures and rows on this page are synthetic and shown to illustrate the format.
- 262
- hours
- 40,000
- utterances
- 900
- contributors
- 7
- countries
- 29.9 GB
- total size
Hours by language
- Indonesian64.5
- Vietnamese37.0
- Hindi28.2
- Javanese26.2
- Thai25.3
- Tagalog23.4
- Bengali16.6
- Sundanese14.5
- Malay14.2
- Cebuano11.8
Hours by setting
- home88.8
- phone call46.8
- street46.7
- workplace42.2
- market37.2
Hours by speaker age
- 18-2998.1
- 30-4489.5
- 45-5952.9
- 60+21.3
Metadata columns
One row per sample, stored as Parquet. Files are FLAC.
| Column | Type | What it holds |
|---|
| utterance_id | Utf8 | Stable identifier for the utterance. |
| speaker_id | Utf8 | Pseudonymous speaker identifier. Split train and test on it. |
| country | Utf8 | Country where the speaker lives. |
| language | Utf8 | Language spoken. |
| duration_s | Float32 | Utterance length in seconds. |
| sample_rate_hz | Int32 | Audio sample rate. |
| setting | Utf8 | home, street, market, workplace or phone_call. |
| speaker_age_band | Utf8 | Speaker's age band. |
| has_transcript | Boolean | Whether a human transcript ships with the utterance. |
Sample rows
| audio_uri | consent_version | country | duration_s | has_transcript | language | sample_rate_hz | setting | size_bytes | speaker_age_band | speaker_id | utterance_id |
|---|
| s3://datachad-lake/datasets/conversational-speech/audio/indonesian/utt_a7cf0399.flac | contributor-consent-v1 | Indonesia | 2.34 | true | Indonesian | 48000 | market | 123585 | 18-29 | spk_0216 | utt_a7cf0399 |
| s3://datachad-lake/datasets/conversational-speech/audio/vietnamese/utt_0f82f009.flac | contributor-consent-v1 | Vietnam | 26.15 | true | Vietnamese | 48000 | street | 1380756 | 45-59 | spk_0455 | utt_0f82f009 |
| s3://datachad-lake/datasets/conversational-speech/audio/javanese/utt_5301d4a9.flac | contributor-consent-v1 | Indonesia | 41.01 | true | Javanese | 16000 |
Read the metadata
The metadata is plain Parquet. Filter it with PyArrow, DuckDB or Polars, then load only the files you picked.
import pyarrow.compute as pc
import pyarrow.dataset as ds
metadata = ds.dataset(
"s3://datachad-lake/datasets/conversational-speech/metadata/",
format="parquet",
)
rows = metadata.to_table(filter=pc.field("country") == "Indonesia")
print(rows.num_rows, rows.column_names)