hysi-lab/RadioTalk-Reproduction-20261002

RadioTalk-style Reproduction — VOA Learning English

Newly acquired English educational broadcast archives, with original publisher MP3 bytes, derived 16 kHz mono PCM WAV clips, automatic transcripts, word timestamps and per-episode speaker IDs. Collection was initialized on 2026-10-02; source publication dates are preserved separately. The initial pilot recordings were published in 2025; the historical expansion includes older source publication dates.

This is an independent corpus made using the processing approach in mit-ccc/audio-scraper, from the group behind RadioTalk. It does not contain recovered original 2018–2019 RadioTalk recordings, and is not an exact replication of its Kaldi-era system, station sampling or talk-radio distribution.

Sources and coverage

Historical acquisition follows every native-audio RSS entry in the publisher's podcast directory, including all reachable article and episode archive pagination. The initial directory exposed 17 audio programs. acquisition_scope.json is the authoritative expanded source catalog; the two sources entries in the frozen processing_config.json are the original pilot seeds, retained to preserve processing-contract identity.

There is no total item or page cap. coverage.json reports fetched, pending, failed and explicitly exhausted archive pages, saved response hashes, and source coverage. A running crawl is not a claim of complete historical coverage. status.json records measured original-audio hours, processing states and publication totals. Original RSS snapshots are retained as seed provenance; the initial 500-source sample is reused and deduplicated against historical discovery. Small transfer and upload batches continuously refill until discovery and the accepted processing queues are exhausted.

The scope is native MP3 audio, not video-only podcasts. Unavailable pages/files remain in the failure ledger. Programs identified as third-party productions, including What It Takes interviews supplied by the American Academy of Achievement, are enumerated and retained locally for source review instead of being published automatically. Explicit agency credits and automatic processing quality flags are also held and reported in exclusions.jsonl.

Each accepted episode retains its source page, media URL, source publication date when available, original byte count, SHA-256 and ffprobe duration. A missing historical publication date remains null. RSS enclosure sizes are advisory and can be stale; mismatches are explicitly recorded. Successful downloads require transport completeness and full ffmpeg decoding. Credit: Voice of America, learningenglish.voanews.com. The publisher's reuse statement permits educational and commercial reuse with attribution and excludes credited third-party agency material. The collector screens available article text or audio player metadata for explicit AP/Reuters/AFP credits and records the policy and page hashes. Audio player pages must reference the same publisher media ID; the player can advertise a lower bitrate rendition than the original/HQ RSS enclosure that is retained. The review scope is recorded per item; many player pages do not contain a transcript. This automated screen is not an exhaustive rights review of every quoted example or incidental sound. No source page photographs are included in the packages. MIT licensing applies only to the upstream software, not to the entire audio collection.

Processing

The pinned upstream Transcriber.process implements ASR → forced word alignment → diarization → word-speaker assignment. The adaptation uses:

• Whisper large-v3 through faster-whisper and WhisperX 3.8.6; int8/float16 inference.
• The English torchaudio WAV2VEC2_ASR_BASE_960H alignment model; downloaded checkpoint hashes are retained. Original ASR spans are kept during alignment instead of NLTK sentence regrouping; the character/word alignment computation is retained.
• Public Silero VAD plus SpeechBrain ECAPA embeddings and cosine agglomerative clustering (threshold 0.65, 3-second windows, 1.5-second hop). Short voiced spans use centered real audio context for embeddings. This replaces the gated pyannote diarizer. No pyannote gated checkpoints were used. The threshold was selected from a stable clustering interval in two pilot episodes, without gold speaker labels or a measured diarization error rate.
• An episode is processed on its full timeline, then cut near word boundaries at approximately 30 seconds. Original MP3s remain unchanged. Silence and music can remain in derived clips; empty-word clips are omitted.

Speaker IDs are anonymous and local to an episode. They do not establish identities across recordings. Clustering may merge distinct people or split one person; short turns are especially uncertain. There is no overlapping-speech separation, overlap-aware diarization, or human correction. These recordings include instruction, narration and examples, and are not a verified natural two-person conversational dataset.

All annotations carry automatic_unreviewed. Alignment and assignment coverage are mechanical checks, not measured transcription or diarization accuracy. Untimed words remain in the full transcript and counts but are omitted from timed clip text. Episodes below 95% word timing or speaker assignment coverage, or with more than 12 estimated speakers, are held for review.

Files and loading

episodes/<prefix>/<source_id>.tar contains:

• original.mp3: exact bytes downloaded from the publisher.
• chunks/*.wav: 16 kHz, mono, 16-bit PCM clips.
• segments.jsonl: text, word timestamps relative to each clip, source offsets, speaker IDs, waveform hashes and archive/member paths.
• source.json, processing.json, quality.json, full_transcript.json, diarization.json, member_checksums.json.

records/<prefix>/<source_id>.jsonl provides the same segment index without downloading the audio archives. The default Hugging Face JSON configuration loads metadata; audio is inside the referenced tar member, rather than embedded in the JSON or automatically playable by the dataset viewer. load_episode.py returns verified WAV bytes without extracting arbitrary archive paths.

For a direct preview, open samples/original.mp3 or samples/processed.wav. These duplicate one included episode and clip for convenience and are not counted again in corpus totals.

from huggingface_hub import HfApi, hf_hub_download
import importlib.util

repo = "hysi-lab/RadioTalk-Reproduction-20261002"
revision = HfApi().repo_info(repo, repo_type="dataset").sha
files = HfApi().list_repo_files(repo, repo_type="dataset", revision=revision)
source_id = next(p for p in files if p.startswith("episodes/") and p.endswith(".tar")).split("/")[-1][:-4]
path = hf_hub_download(repo, "load_episode.py", repo_type="dataset", revision=revision)
spec = importlib.util.spec_from_file_location("episode_reader", path)
reader = importlib.util.module_from_spec(spec)
spec.loader.exec_module(reader)
for row in reader.iter_episode(source_id, revision=revision):
    print(row["text"], len(row["audio_bytes"]))

Each package is checked against its local SHA-256 and the Hub's LFS digest after upload. The first package is additionally downloaded back in full and hashed. Exact duplicate source bytes are excluded within this acquisition; near-duplicates and train/test leakage have not been audited. No evaluation split or demographic labels are provided.

Continuous acquisition workers

5080 performs full archive discovery, local acquisition, GPU processing and final publication. A800 concurrently downloads, decodes and checks disjoint original-audio sources; its actual role is acquisition, not GPU inference. workers/a800-gpu7/full_assignment.json contains the continuously replenished transfer assignment. Originals under incoming/a800-gpu7/ are verified acquisition checkpoints; they are excluded from the default dataset config and corpus totals. The controller independently verifies their size, SHA-256 and full decoding before processing. A800 only releases its expendable working copy after controller acknowledgement; the controller's original and the HF checkpoint remain retained.