chika / README.md
internetoftim's picture
Register gated curation snapshot hil_dwnd_batch_asr_6127_v1
4a4bed9 verified
|
Raw History Blame Contribute Delete
23.8 kB
metadata
language:
  - fil
  - tgl
  - hil
  - eng
  - ceb
  - bik
  - ilo
license: other
license_name: source-specific-rights-unresolved
pretty_name: Chika — Philippine Languages Speech Curation
multilinguality:
  - multilingual
task_categories:
  - automatic-speech-recognition
  - text-to-speech
tags:
  - philippines
  - taglish
  - hiligaynon
  - code-switching
  - work-in-progress
configs:
  - config_name: hil_dwnd_batch_asr_6127_v1
    data_files:
      - split: unassigned
        path: data/hil_dwnd_batch_asr_6127_v1/curation-unassigned-*.parquet
  - config_name: ilo_dwnd_batch_asr_29398_v1
    data_files:
      - split: unassigned
        path: data/ilo_dwnd_batch_asr_29398_v1/curation-unassigned-*.parquet
  - config_name: ceb_dwnd_batch_asr_32311_v1
    data_files:
      - split: unassigned
        path: data/ceb_dwnd_batch_asr_32311_v1/curation-unassigned-*.parquet
  - config_name: hil_dwnd_batch_asr_59170_v1
    data_files:
      - split: unassigned
        path: data/hil_dwnd_batch_asr_59170_v1/curation-unassigned-*.parquet
  - config_name: bik_dwnd_batch_asr_40574_v1
    data_files:
      - split: unassigned
        path: data/bik_dwnd_batch_asr_40574_v1/curation-unassigned-*.parquet
  - config_name: fil_dwnd_batch_asr_62826_v1
    data_files:
      - split: unassigned
        path: data/fil_dwnd_batch_asr_62826_v1/curation-unassigned-*.parquet
  - config_name: hil_dwnd_batch_asr_54272_v1
    data_files:
      - split: unassigned
        path: data/hil_dwnd_batch_asr_54272_v1/curation-unassigned-*.parquet
  - config_name: fil_dwnd_asr_9927_v1
    data_files:
      - split: unassigned
        path: data/fil_dwnd_asr_9927_v1/curation-unassigned-*.parquet
  - config_name: hil_dwnd_asr_1181_v1
    data_files:
      - split: unassigned
        path: data/hil_dwnd_asr_1181_v1/curation-unassigned-*.parquet
  - config_name: fil_dwnd_asr_4000_v1
    data_files:
      - split: unassigned
        path: data/fil_dwnd_asr_4000_v1/curation-unassigned-*.parquet
  - config_name: fil_dwnd_asr_2000_v1
    data_files:
      - split: unassigned
        path: data/fil_dwnd_asr_2000_v1/curation-unassigned-*.parquet
  - config_name: fil_collected_vad_5671_v1
    data_files:
      - split: unassigned
        path: data/fil_collected_vad_5671_v1/curation-unassigned-*.parquet
  - config_name: fil_collected_vad_982_v1
    data_files:
      - split: unassigned
        path: curation/collected-fil-982-v1/curation-unassigned-*.parquet
  - config_name: hil_collected_vad_718_v1
    data_files:
      - split: unassigned
        path: curation/collected-hil-718-v1/curation-unassigned-*.parquet
  - config_name: hil_collected_vad_667_v2
    data_files:
      - split: unassigned
        path: curation/collected-hil-667-v2/curation-unassigned-*.parquet
  - config_name: fil_collected_vad_482_v1
    data_files:
      - split: unassigned
        path: curation/collected-fil-482-v1/curation-unassigned-*.parquet
  - config_name: hil_collected_vad_667_v1
    data_files:
      - split: unassigned
        path: curation/collected-hil-667-v1/curation-unassigned-*.parquet
  - config_name: fil_collected_vad_100_v1
    data_files:
      - split: unassigned
        path: curation/collected-fil-vad-100-v1/curation-unassigned-*.parquet
  - config_name: hil_collected_vad_200_reviewed_v2
    data_files:
      - split: unassigned
        path: >-
          curation/collected-hil-vad-200-reviewed-v2/curation-unassigned-*.parquet
  - config_name: hil_collected_vad_200_v1
    data_files:
      - split: unassigned
        path: curation/collected-hil-vad-200-v1/curation-unassigned-*.parquet
  - config_name: hil_collected_pilot_v1
    data_files:
      - split: unassigned
        path: curation/collected-hil-pilot-v1/curation-unassigned-*.parquet
  - config_name: fil_collected_pilot_v1
    data_files:
      - split: unassigned
        path: curation/collected-fil-pilot-v1/curation-unassigned-*.parquet
  - config_name: hil_pld_curation_v2
    data_files:
      - split: unassigned
        path: curation/pld-hil-v2/curation-unassigned-*.parquet
      - split: reserved_test
        path: curation/pld-hil-v2/curation-reserved_test-*.parquet
  - config_name: hil_pld_curation_v1
    data_files:
      - split: unassigned
        path: curation/pld-hil-v1/curation-unassigned-*.parquet
      - split: reserved_test
        path: curation/pld-hil-v1/curation-reserved_test-*.parquet

Chika

A work-in-progress collection of Filipino/Taglish, Hiligaynon, Cebuano, Bikolano and Ilocano speech, preserving natural English code-switching. The target is approximately 250 unique retained, transcribed hours per language group (1,250 hours overall). This target is not a claim about currently completed data.

Tracking and workflow: sapinsapin/halohalo issue #8. The repository is private while source rights and curation are unresolved.

Available configuration

hil_dwnd_batch_asr_6127_v1 contains the remaining 6,127 Hiligaynon-source clips (44,519.24 padded seconds, about 12.4 hours) of offline batch run remote-hil-asr-local-v1: the rows of its fully audited 236-chunk output that are not in hil_dwnd_batch_asr_54272_v1 (the run's first 212 chunks). Aksyon and GMA Ratsada broadcast recordings, draft ASR outputs with unassigned splits and no empty transcripts, from sapinsapin/whisper-large-v3-pld-hil-norm at revision 7a695b0f737b3e7cec482c29f0d12c2126cb4f4c; 370 rows used temperature fallback and 8 rows have an unresolved repetition flag in provenance_json. Together with hil_dwnd_batch_asr_54272_v1 it covers the whole run. It shares no clip IDs or audio hashes with hil_dwnd_asr_1181_v1, hil_dwnd_batch_asr_54272_v1 or hil_dwnd_batch_asr_59170_v1. Source language is provisional; speakers and rights remain unknown. Curation snapshot, not approved training data.

ilo_dwnd_batch_asr_29398_v1 contains 29,398 Ilocano-source clips (256,635.632 padded seconds, about 71.3 hours) from 169 ABS-CBN TV Patrol Ilocos recordings, all draft ASR outputs with unassigned splits and no empty transcripts. Text comes from offline batch ASR on dwnd-01 with sapinsapin/whisper-large-v3-pld-ilo-norm at revision 00f3c4812849d7abfc0bb3519741b466457368ab; 769 rows used temperature fallback and 5 rows have an unresolved repetition flag in provenance_json. This model is trained on read PLD speech; on a 50-clip broadcast sample it disagreed with a provisional Gemini reference at about 47% WER, so treat transcripts as rough drafts. Clips are VAD windows of at most 27 s. Every row was audited against the frozen batch plan, ASR configuration and re-hashed source audio. Source language is provisional; speakers and rights remain unknown. This is a curation snapshot, not approved training data.

ceb_dwnd_batch_asr_32311_v1 contains 32,311 Cebuano-source clips (258,222.59 padded seconds, about 71.7 hours) from 165 GMA Regional TV Balitang Bisdak recordings, all draft ASR outputs with unassigned splits and no empty transcripts. Text comes from offline batch ASR on dwnd-01 with sapinsapin/whisper-large-v3-pld-ceb-norm at revision 109c7f42e40bfa071d53270115734dc2d88fb521; 1,247 rows used temperature fallback and 145 rows have an unresolved repetition flag in provenance_json. This model is trained on read PLD speech; on a 50-clip broadcast sample it disagreed with a provisional Gemini reference at about 44% WER, so treat transcripts as rough drafts. Clips are VAD windows of at most 27 s. Every row was audited against the frozen batch plan, ASR configuration and re-hashed source audio. Source language is provisional; speakers and rights remain unknown. This is a curation snapshot, not approved training data.

hil_dwnd_batch_asr_59170_v1 contains 59,170 Hiligaynon-source clips (339,059.95 padded seconds, about 94.2 hours) from 100 Aksyon stream and GMA Ratsada broadcast recordings, all draft ASR outputs with unassigned splits and no empty transcripts. Text comes from offline batch ASR on dwnd-01 with sapinsapin/whisper-large-v3-pld-hil-norm at revision 7a695b0f737b3e7cec482c29f0d12c2126cb4f4c; 2,156 rows used temperature fallback and 66 rows have an unresolved repetition flag in provenance_json. Clips are VAD windows of at most 27 s. Every row was audited against the frozen batch plan, ASR configuration and re-hashed source audio. It shares no clip IDs or audio hashes with hil_dwnd_asr_1181_v1 or hil_dwnd_batch_asr_54272_v1. Source language is provisional; speakers and rights remain unknown. This is a curation snapshot, not approved training data; deduplicate by provenance audio hashes before combining configurations.

bik_dwnd_batch_asr_40574_v1 contains 40,574 Bikolano-source clips (280,065.184 padded seconds, about 77.80 hours, 0 empty transcripts) from 165 GMA Regional TV Balitang Bicolandia recordings (remote-bik-vad-full-v1 and remote-bik-processing-v2..v4), all draft ASR outputs with unassigned splits and a nullable translation_en column (present in schema, null until translated). Text comes from offline batch ASR on dwnd-01 with sapinsapin/whisper-large-v3-pld-bcl-norm at revision a35847d1db75e5c72409be4cc7f8b03775b45fb7; 219 rows used temperature fallback and 8 rows have an unresolved repetition flag in provenance_json. Clips are VAD windows of at most 27 s. Every included row was audited on dwnd-01 against the frozen plan, ASR configuration and re-hashed source audio, and exported into 41 Parquet shards. Source language and Bikol variety are provisional; speakers and rights remain unknown. This is a gated curation snapshot, not approved training data; deduplicate by provenance audio hashes before combining with any configuration.

fil_dwnd_batch_asr_62826_v1 contains 62,826 Filipino/Taglish clips (279,082.961 padded seconds, about 77.52 hours) from the frozen remaining Filipino VAD backlog (remote-fil-asr-local-v1), all draft ASR outputs with unassigned splits and a nullable translation_en column (present in schema, null until translated). Text comes from offline batch ASR on reth5et (RTX 3070) with sapinsapin/whisper-large-v3-pld-fil-norm at revision ee5b4f0cd1af616b38f742ae05b644d6783b52fe; 462 rows used temperature fallback and 1 row has an unresolved repetition flag in provenance_json. Clips are VAD windows of at most 27 s. Every included row was audited on dwnd-01 against the frozen plan, ASR configuration and re-hashed source audio, and exported into 63 Parquet shards. Source language is provisional; speakers and rights remain unknown. This is a gated curation snapshot, not approved training data. It shares no clip IDs or audio hashes with fil_dwnd_asr_1000 (first1000), fil_dwnd_asr_2000_v1, fil_dwnd_asr_4000_v1, or fil_dwnd_asr_9927_v1; deduplicate by provenance audio hashes before combining with any configuration.

hil_dwnd_batch_asr_54272_v1 contains 54,272 Hiligaynon-source clips (333,508.521 padded seconds, about 92.6 hours) from Aksyon streams and videos and GMA Ratsada broadcast recordings, all draft ASR outputs with unassigned splits. Text comes from offline batch ASR on dwnd-01 with sapinsapin/whisper-large-v3-pld-hil-norm at revision 7a695b0f737b3e7cec482c29f0d12c2126cb4f4c; 2,330 rows used temperature fallback and 49 rows have an unresolved repetition flag in provenance_json. Clips are VAD windows of at most 27 s. The batch run was stopped after 212 of 236 planned chunks; the 6,127 clips in unfinished chunks are not included and remain untranscribed. Every included row was audited against the frozen plan, ASR configuration and re-hashed source audio. Source language is provisional; speakers and rights remain unknown. This is a private curation snapshot, not approved training data. It shares no clip IDs or audio hashes with hil_dwnd_asr_1181_v1; deduplicate by provenance audio hashes before combining with any configuration.

fil_dwnd_asr_9927_v1 contains 9,927 Filipino/Taglish-source clips (44,233.244 padded seconds), all draft ASR outputs with unassigned splits, including 46 empty transcripts. It comes from remote ASR queue v6; one request that hit the server finalize deadline is excluded and quarantined, and the 72 later clips in that batch were not transcribed. It shares no clip ID or audio hash with fil_dwnd_asr_2000_v1 or fil_dwnd_asr_4000_v1. Source language is provisional; speakers and rights remain unknown. All exported PCM samples were checked against original audio. This is a private curation snapshot, not approved training data. Deduplicate by provenance audio hashes before combining configurations.

hil_dwnd_asr_1181_v1 contains 1,181 Hiligaynon-source clips (5,313.585 padded seconds) from Aksyon and GMA Ratsada broadcast recordings, all draft ASR outputs with unassigned splits, including one empty transcript. Clips are at most 12 s except one row from the first bounded queue. Forty requests that hit the server finalize deadline are excluded and quarantined. Source language is provisional; speakers and rights remain unknown. All exported PCM samples were checked against original audio. This is a private curation snapshot, not approved training data. It shares no audio with earlier Hiligaynon configurations, which come from different sources; deduplicate by provenance audio hashes before combining.

fil_dwnd_asr_4000_v1 contains 4,000 Filipino/Taglish-source clips (17,245.148 padded seconds), all draft ASR outputs with unassigned splits, including fifteen empty transcripts. It combines remote ASR queues v4 and v5 and shares no clip ID or audio hash with fil_dwnd_asr_2000_v1. Source language is provisional; speakers and rights remain unknown. All exported PCM samples were checked against original audio. This is a private curation snapshot, not approved training data. Deduplicate by provenance audio hashes before combining configurations.

fil_dwnd_asr_2000_v1 contains 2,000 Filipino/Taglish-source clips (8,186.89 padded seconds), all draft ASR outputs with unassigned splits, including ten empty transcripts. Source language is provisional; speakers and rights remain unknown. All exported PCM samples were checked against original audio. This is a private curation snapshot, not approved training data. It may overlap other snapshots; deduplicate by provenance audio hashes before combining configurations. Text reviews performed later are not included in this immutable snapshot.

fil_collected_vad_5671_v1 contains 5,671 Filipino/Taglish-source clips (25,099.226375 padded seconds), with 1,061 validated text-only reviews and 4,610 drafts, including 19 empty transcripts. It combines completed ASR v4 with the earlier 982-clip snapshot and 100 Kimi reviews. This overlaps older configurations; do not sum or concatenate them as independent recordings.

All 5,671 exported PCM arrays and transcripts were checked against local inputs. Source language remains provisional, rights unknown, splits unassigned and audio approvals zero. Kimi requested route us.kimi-k3 returned global.moonshotai.kimi-k3.1; earlier reviews retain their original model provenance. Text review is not audio verification. This is a curation snapshot, not an approved training split.

fil_collected_vad_982_v1 contains 982 Filipino/Taglish-source segments (3,730.272 padded seconds), with 961 validated text-only reviews and 21 draft-only rows. It includes the earlier 482-row snapshot plus 500 additional nonempty ASR drafts, of which 490 reviews passed validation and ten were rejected. No rejected review was repaired or attached. Do not sum overlapping configuration durations.

Audio and transcripts were preserved; all splits remain unassigned, source rights unknown, and audio approvals zero. Exact PCM and unchanged sentences were checked for all 982 rows; remote export artifact hashes were verified. These are curation drafts, not an approved training split.

hil_collected_vad_718_v1 contains 718 Hiligaynon-source segments (1,892.95 padded seconds), with 705 validated text-only reviews and 13 draft-only rows. It includes the 667-row snapshots plus 51 newly transcribed and reviewed segments from different original recordings. The initial ten-segment persistent-session pilot is already included. Do not sum overlapping configuration durations. The ASR continuation stopped on a server finalization timeout; no failed request is included.

All transcripts and audio were preserved, all splits remain unassigned, rights remain unknown, and audio approvals are zero. Every embedded PCM array and sentence was checked against the local manifest; remote artifact hashes were verified. This is a curation draft, not a training split.

hil_collected_vad_667_v2 preserves the same 667 audio segments and transcripts as v1 (1,596.56 padded seconds), with 654 validated text-only reviews and 13 draft-only rows whose reviews were rejected. It adds review metadata, not new audio. All reviews were attempted; audio approvals remain zero.

fil_collected_vad_482_v1 contains 482 Filipino/Taglish-source segments (1,924.68 padded seconds), with 471 validated text-only reviews and 11 draft-only rows whose reviews were rejected. It includes the earlier 100-row Filipino configuration. Audio and transcript text were preserved.

Both snapshots are immutable curation drafts: all rows are unassigned, rights remain unknown, and no audio approval is implied. Exact embedded PCM and unchanged sentences were checked against their local manifests; published artifact hashes were verified. Do not concatenate overlapping configurations or sum their durations as unique hours.

hil_collected_vad_667_v1 contains all 667 successful draft transcripts from the bounded Hiligaynon batch: 1,596.56 padded seconds (26.61 minutes). There are 531 validated text-only reviews and 136 draft-only rows (11 rejected reviews, 125 not attempted at snapshot time). All transcripts and audio are unchanged; zero audio approvals. This includes the earlier 200-row Hiligaynon snapshots, so do not concatenate configurations or add their durations as unique audio. All splits remain unassigned and rights unknown. Later review work does not mutate this frozen snapshot.

fil_collected_vad_100_v1 contains 100 Filipino/Taglish-source VAD candidates, 292.476 padded seconds (about 4.9 minutes), with ASR drafts and 99 validated text-only reviews. One review failed and remains unattached. Audio and transcripts were preserved, all splits are unassigned, rights remain unknown and audio approvals are zero. This overlaps earlier Filipino pilot source intervals; do not add configuration durations as unique hours.

hil_collected_vad_200_reviewed_v2 supersedes the review metadata of the 200-row v1 snapshot: 194 validated text-only LLM reviews and six rows without a successful review. Audio and transcripts are unchanged (489.686 padded seconds); zero audio approvals are granted. This is the same audio as hil_collected_vad_200_v1, not additional unique hours. All source-rights and split restrictions remain unchanged.

hil_collected_vad_200_v1 contains 200 Hiligaynon-source VAD candidate segments, 489.686 seconds (about 8.2 minutes), with ASR drafts. Five have validated text-only LLM reviews; 195 remain draft-only. All are unassigned with unknown source permissions and zero audio verification. Padded VAD candidates are not certified clean speech. This snapshot overlaps the earlier collected pilot sources; do not sum configuration durations or concatenate them for training without source-interval/acoustic deduplication.

Cebuano, Bikolano and Ilocano acquisition is being added; no completed transcribed configurations for those languages are claimed. Bikolano is a collection label: specific varieties and unresolved variety labels remain explicit in source metadata.

hil_collected_pilot_v1 contains 3 newly collected transport windows (90 seconds), one with a validated text-only LLM review. fil_collected_pilot_v1 contains 2 windows (60 seconds). All are unassigned, machine-transcribed drafts, with source-level language labels and unknown training/commercial/redistribution rights. These are integration pilots, not clean speech or accuracy benchmarks. Automatic ASR cuts can split phrases within a transport window; music, ads, mixed speakers and English-only content may remain. Bulk ASR was stopped after server finalization timeouts. No audio verification is claimed.

hil_pld_curation_v2 contains 229 PLD spontaneous recordings, 2.9522 hours, with machine-generated transcript drafts and text-only review. There are 190 unassigned and 39 reserved-test recordings. These recordings are an initial schema migration/curation set; they do not establish completion of the new 250-hour Hiligaynon collection. No audio-verified training labels are claimed. PLD source restrictions remain applicable.

from datasets import load_dataset
rows = load_dataset("sapinsapin/chika", "hil_pld_curation_v2", split="unassigned", streaming=True)
row = next(iter(rows))
print(row["sentence"], row["review_status"])

raw/ contains source audio backups; checkpoints/ contains acquisition ledgers. Neither is an approved training split. Runtime totals in acquisition ledgers include music, ads and silence. Feed estimates, downloaded hours, retained speech, ASR-completed hours, LLM-reviewed hours and human/audio-verified hours must be reported separately. Simulcasts, excerpts, reposts and alternate encodings require acoustic deduplication before unique-hour accounting.

Format

Following PLD, each Parquet row embeds audio as {bytes, path} with 16 kHz mono FLAC, plus sentence, duration, speaker_id, speech_type, language and num_words. The prompt distinction is directly available as text_is_prompt and prompt, with source_id and recording_id for grouping. Other added fields are id, split, schema_version, review_status and provenance_json. A complete manifest sidecar preserves source questions, source hashes, original timestamps, review evidence, rights and original sample rate.

For spontaneous PLD speech, sentence is the draft transcription, never the question. The elicitation question remains separate in provenance and is excluded from ASR and text-review inputs. Frozen test reservations must not be used for training. Source recording and known speaker grouping governs future splits; unknown speakers do not imply speaker-disjointness.

The older hil_pld_curation_v1 configuration remains reproducible but keeps the prompt flag inside provenance_json; use v2 for direct compatibility with PLD’s not row["text_is_prompt"] filter.

Language and review

Filipino–English and Hiligaynon–English code-switching are valid speech, including local affixes on English roots. No language purification or translation is applied. Source labels are provisional until the audio is checked. An English title or feed-level language tag does not determine the recording's spoken language.

LLM review checks textual plausibility and flags uncertainty. It does not prove fidelity to audio, approve a correction, or provide a ground-truth WER/CER reference. Raw drafts and suggestions remain distinguishable. TTS use additionally requires speaker consistency, speech-only segments and accurate text/audio alignment.

Sources and rights

Sources include public podcast enclosures, regional news/video archives, bounded radio captures, and the existing PLD curation pilot. Exact source URLs, retrieval metadata, hashes and restrictions remain attached to each record. Public download availability does not grant commercial training or redistribution rights. There is no blanket permissive license for this collection. Unknown rights remain unknown; restricted sources must be excluded from any incompatible release or training profile. Private storage does not resolve these rights.

Synthetic audio, if later included, will have explicit model/version/license provenance and will be counted separately from unique real recorded hours. No synthetic hours currently contribute to the acquisition target.