pew-data-labs/religious-radio-corpus

Religious Radio Corpus

A large-scale corpus of transcribed English-language religious radio broadcasts, captured from live webstreams in the United States over a one-month window in July 2025. Fifteen-minute segments were recorded on a rolling schedule from 779 distinct streams, which together rebroadcast the signals of more than 2,000 AM and FM stations, yielding 715,688 recordings and over 60 million diarized lines of speech. Each recording was transcribed and speaker-diarized with an automated pipeline, then segmented and labeled by programming format and topic using a large language model.

Released corpus contains only derived transcripts and annotations, the raw audio recordings are not included.

Full details on data collection, methods, and validation are in the accompanying paper: A large-scale corpus of religious radio broadcast transcripts from webstream recordings in the United States (arXiv: https://doi.org/10.48550/arXiv.2607.26249).

Dataset Structure

The corpus is organized as three linked tables — stations, recordings, and transcript_lines — plus a crosswalk mapping each monitored stream to the additional stations that carry its signal.

stations/stations.parquet
recordings/recordings.parquet
crosswalk/station_stream_crosswalk.parquet
transcript_lines/YYYY-MM-DD/part-*.parquet
README.md

The transcript_lines table is partitioned by broadcast date (UTC), with one subdirectory per day from 2025-07-01 through 2025-08-01, so a single day or range can be loaded without reading the full corpus.

Tables join on stable identifiers: recordings.station_id → stations.station_id, and transcript_lines.recording_id → recordings.recording_id. The crosswalk expands stream-level content to the station level.

stations

One row per monitored stream URL (the operational recording unit); 779 rows.

Field	Type	Description
station_id	string	Primary key; unique identifier for the stream entry
callsign	string	Call sign of the station associated with the stream (e.g., WAMU-FM)
band	string	Frequency band: AM or FM
region	string	U.S. Census region: Northeast, South, Midwest, or West
timezone	string	Station's local timezone in IANA format (e.g., America/Chicago)
utc_adjust	float	Offset in hours between the station's local time and UTC
related_stations	list[string]	Identifiers of all stations that share this stream URL

station_stream_crosswalk

One row per station–stream pair.

Field	Type	Description
station_id	string	Foreign key to stations.station_id (the monitored stream entry)
stream_station_id	string	Identifier of an individual station carried by the stream
is_primary	boolean	True for primary stations (AM, FM, low-power FM); false for boosters and translators

recordings

One row per captured 15-minute recording.

Field	Type	Description
recording_id	integer	Primary key; unique identifier for the recording
station_id	string	Foreign key to stations.station_id
day	string	Broadcast date of the recording (YYYY-MM-DD)
recording_start_utc	timestamp	Start time of the recording in UTC (ISO 8601)
recording_duration_sec	float	Duration of the recording in seconds
predicted_programming_formats	list[string]	Distinct programming-format labels across the recording's segments
predicted_topic_labels	list[string]	Distinct topic labels across the recording's segments

The two label fields hold the union of labels assigned to programming segments within a given recording; predicted_topic_labels may be empty.

transcript_lines

One row per diarized utterance (or music region).

Field	Type	Description
recording_id	integer	Foreign key to recordings.recording_id
station_id	string	Identifier of the associated station (denormalized for convenience)
line_index	integer	Sequential index of the line within its recording (0-based)
start_sec	float	Start time of the line, in seconds from the recording start
end_sec	float	End time of the line, in seconds from the recording start
duration_sec	float	Duration of the line in seconds
speaker_id_local	string	Speaker label, unique only within a recording; <MUSIC> marks a music region
text	string	Transcribed text of the utterance, or <MUSIC> for a music region
is_music	boolean	True if the segment was classified as music

Label codebooks

Format (mutually exclusive): Ad/promotion · Audio drama/narrative · Caller interaction/audience participation · Discussion/monologue/commentary · Interview · News read/traffic/weather · Religious service/sermon · Transition/filler

Topic (multi-label): Business/economics/finance · Entertainment/pop culture/sports · Family/parenting/education · Health/wellness · Lifestyle/advice/personal development · Politics/current events/social commentary · Religion · Science/technology

Usage

Distributed as Apache Parquet and readable with any Parquet-aware tool. transcript_lines is large, users may wish to stream it or load it in parts.

from datasets import load_dataset
 
stations = load_dataset("pew-data-labs/religious-radio-corpus",
                        data_files="stations/stations.parquet", split="train")
 
recordings = load_dataset("pew-data-labs/religious-radio-corpus",
                          data_files="recordings/recordings.parquet", split="train")
 
# A single day of transcript lines (recommended)
day = load_dataset("pew-data-labs/religious-radio-corpus",
                   data_files="transcript_lines/2025-07-01/*.parquet",
                   split="train", streaming=True)
 
# Full transcript corpus via streaming
full = load_dataset("pew-data-labs/religious-radio-corpus",
                    data_files="transcript_lines/**/*.parquet",
                    split="train", streaming=True)

Things to know

• Stream vs. station attribution. Each stream/recording corresponds to a monitored stream URL corresponding to a single distinct station_id, but stations under common ownership often share a feed. Station-level analysis should expand each stream via related_stations or the crosswalk, which attributes the same content to every station on that feed. The corpus is ~60M transcript lines at the stream level, and expands to ~188M lines at the station level. Confirm whether a count refers to streams or stations before comparing figures.
• Speaker IDs (SPK_00, SPK_01, …) are assigned independently per recording and are not stable across recordings. Speaker-level analysis is well defined only within a single recording.
• Music segments are not transcribed; they carry <MUSIC> in text and speaker_id_local and are flagged by is_music. Filter on is_music to keep or drop them.
• Time zones. recording_start_utc is UTC; convert to local broadcast time with utc_adjust or the IANA timezone field for time-of-day or day-of-week analysis.
• Derived fields. Transcripts are ASR/diarization output and the format/topic labels are LLM-generated; both carry error that varies by content type and audio quality. See the paper's Technical Validation section for measured error rates (ASR WER ~5% overall; format/topic macro-F1 of 0.77/0.71).

License

This dataset (the "Religious Radio Corpus") is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. The full license text is available at https://creativecommons.org/licenses/by/4.0/. The license applies only to rights, if any, held by the dataset creators in the scope identified below and does not grant any rights in underlying radio broadcasts or other third-party content in or derived from those broadcasts.

Rights and Limitations

Scope of the license

The CC BY 4.0 license applies to the components created by the dataset creators, including:

• The dataset structure and organization
• Creator-generated elements associated with the machine-generated transcripts
• Annotations, metadata, labels, and other creator-generated derived features (e.g., topic and format labels)

To the extent the dataset creators hold copyright or other intellectual property rights in these components, such rights are licensed under CC BY 4.0. The dataset creators do not claim ownership of, and do not grant rights in, copyrightable expression originating in underlying broadcasts or other third party materials.

Third-party content

This dataset contains transcripts derived from radio broadcasts. These broadcasts may include copyrighted, trademarked or otherwise protected material, such as:

• News programming
• Advertisements
• Music and other audio content
• Talk programming, scripts, performances, sound recordings, trademarks, logos, and branding

The CC BY 4.0 license applies only to rights, if any, held by the dataset creators in the Dataset Components and does not apply to the underlying broadcasts or any content contained therein.

No rights are granted with respect to any copyrighted, trademarked, or otherwise protected material originating from the broadcasts.

The inclusion of material derived from a broadcast in this dataset should not be interpreted as permission to reproduce, distribute, publicly perform, publicly display, modify, create derivative works from, train models on, or otherwise use the underlying broadcast content or third party materials contained therein.

All rights in such third-party content remain with their respective owners.

Transcripts and Underlying Expression

Machine-generated transcripts may reflect, reproduce, summarize, or otherwise contain portions of copyrighted expression originating from underlying broadcasts. The dataset creators do not claim ownership of any such underlying expression and do not grant any license to that expression.

To the extent a transcript contains copyrighted expression from an underlying broadcast, any rights in that expression remain with the applicable rights holder.

Source data and station selection

The set of radio stations included in this dataset was compiled using a combination of publicly accessible information and internal selection processes.

This dataset does not reproduce or redistribute any third-party proprietary database, including the Radio-Locator database (Theodric Technologies LLC) referenced in the accompanying documentation. Any station-level fields included in the dataset reflect the dataset creators' own compilation and processing and are not a copy of, or substitute for, any third-party database. Users seeking the underlying commercial or administrative data sources should obtain them directly from their respective providers under those providers' terms.

User responsibility

Users are responsible for ensuring that their use of the dataset complies with applicable legal requirements in their jurisdiction, including, without limitation, copyright, licensing, privacy, publicity, and trademark.

Users are solely responsible for determining whether additional permissions, licenses, or authorizations from third-party rights holders are required for any particular use of material reflected in or derived from the dataset.

Intended Use

This dataset is intended to support research and analysis in areas such as:

• Media and communications research
• Speech and language processing
• Computational social science
• Public discourse analysis

The dataset is not intended to serve as a substitute for access to original broadcast content or as a licensed repository of such content. Nothing in this dataset or accompanying license should be construed as authorizing the redistribution, republication, retransmission, commercial exploitation, or other reuse of broadcast content or third-party materials.

Limitations

• Transcripts are machine-generated and may contain errors from automatic speech recognition (ASR) and speaker diarization.
• Some content (e.g., music, overlapping speech, low-quality streams) may be imperfectly represented.
• Metadata fields such as topics or formats are model-derived and may be noisy or uncertain.

No warranty

The dataset is provided “as is” and “as available,” without warranties of any kind, express or implied, including without limitation warranties of accuracy, completeness, non-infringement, merchantability, fitness for a particular purpose, or suitability for any particular research or other use.

Without limiting the disclaimers in the CC BY 4.0 license, the dataset creators disclaim responsibility for errors, omissions, defects, inaccuracies, transcription errors, metadata errors, model-generated classification errors, or rights-related issues associated with the dataset or underlying materials.

Citation

Please cite the dataset as:

Bestvater, Samuel, Athena Chapekis, Skyler Seets, Anna Lieb, Sono Shah, and Aaron Smith. 2026. “Religious Radio Corpus.” Harvard Dataverse. https://doi.org/10.7910/DVN/NJ71JJ.

Acknowledgements

This work was supported by the Pew-Knight Initiative, a research program jointly funded by The Pew Charitable Trusts and the John S. and James L. Knight Foundation.