Datasets:

knesset-asr
/

knesset-committees-inference

chunk_id large_stringlengths 23 36	speaker_id int64 421 30.9k	session int64 2.07M 2.24M	duration_s float64 1 30	in_subset bool 2 classes	done_A bool 2 classes	done_B bool 2 classes	done_A_auto bool 2 classes	error_A float64	error_B float64
4405_2073683_14600_35160.flac	4,405	2,073,683	20.56	false	false	true	true	null	null
4405_2073683_35160_62600.flac	4,405	2,073,683	27.44	true	true	true	true	null	null
4405_2073683_62600_88240.flac	4,405	2,073,683	25.64	false	false	true	true	null	null
4405_2073683_88240_110860.flac	4,405	2,073,683	22.62	false	false	true	true	null	null
4405_2073683_130220_154860.flac	4,405	2,073,683	24.64	false	false	true	true	null	null
4405_2073683_154860_175120.flac	4,405	2,073,683	20.26	true	true	true	true	null	null
4405_2073683_175120_202320.flac	4,405	2,073,683	27.2	true	true	true	true	null	null
4405_2073683_202320_225340.flac	4,405	2,073,683	23.02	false	false	true	true	null	null
4405_2073683_225340_250780.flac	4,405	2,073,683	25.44	false	false	true	true	null	null
4405_2073683_250780_269740.flac	4,405	2,073,683	18.96	true	true	true	true	null	null
4405_2073683_269740_290480.flac	4,405	2,073,683	20.74	false	false	true	true	null	null
4405_2073683_290480_315780.flac	4,405	2,073,683	25.3	false	false	true	true	null	null
4405_2073683_315780_341100.flac	4,405	2,073,683	25.32	false	false	true	true	null	null
4405_2073683_341100_350900.flac	4,405	2,073,683	9.8	false	false	true	true	null	null
4405_2073683_457200_459420.flac	4,405	2,073,683	2.22	true	true	true	true	null	null
4405_2073683_670750_674100.flac	4,405	2,073,683	3.35	false	false	true	true	null	null
4405_2073683_857960_882440.flac	4,405	2,073,683	24.48	false	false	true	true	null	null
4405_2073683_927960_929180.flac	4,405	2,073,683	1.22	true	true	true	true	null	null
4405_2073683_974680_981140.flac	4,405	2,073,683	6.46	true	true	true	true	null	null
30065_2073683_981540_983410.flac	30,065	2,073,683	1.87	true	true	true	true	null	null
4405_2073683_983440_985680.flac	4,405	2,073,683	2.24	false	false	true	true	null	null
4405_2073683_1089540_1091940.flac	4,405	2,073,683	2.4	true	true	true	true	null	null
4405_2073683_1199320_1203620.flac	4,405	2,073,683	4.3	false	false	true	true	null	null
4405_2073683_1225080_1235260.flac	4,405	2,073,683	10.18	false	false	true	true	null	null
4405_2073683_1254960_1257790.flac	4,405	2,073,683	2.83	false	false	true	true	null	null
30065_2073683_1307960_1313360.flac	30,065	2,073,683	5.4	true	true	true	true	null	null
4405_2073683_1314150_1317840.flac	4,405	2,073,683	3.69	false	false	true	true	null	null
4405_2073683_1322280_1324220.flac	4,405	2,073,683	1.94	false	false	true	true	null	null
4405_2073683_1382860_1384800.flac	4,405	2,073,683	1.94	false	false	true	true	null	null
4405_2073683_1566280_1571060.flac	4,405	2,073,683	4.78	false	false	true	true	null	null
4405_2073683_1579800_1595380.flac	4,405	2,073,683	15.58	false	false	true	true	null	null
4405_2073683_1604300_1607740.flac	4,405	2,073,683	3.44	true	true	true	true	null	null
4405_2073683_1622760_1632700.flac	4,405	2,073,683	9.94	false	false	true	true	null	null
4405_2073683_1648060_1666260.flac	4,405	2,073,683	18.2	false	false	true	true	null	null
30548_2073683_1667680_1693740.flac	30,548	2,073,683	26.06	false	false	true	true	null	null
30548_2073683_1693740_1722120.flac	30,548	2,073,683	28.38	true	true	true	true	null	null
30548_2073683_1722120_1750900.flac	30,548	2,073,683	28.78	true	true	true	true	null	null
30548_2073683_1750900_1771720.flac	30,548	2,073,683	20.82	false	false	true	true	null	null
4405_2073683_1778580_1784300.flac	4,405	2,073,683	5.72	true	true	true	true	null	null
30548_2073683_1784760_1814000.flac	30,548	2,073,683	29.24	false	false	true	true	null	null
30548_2073683_1814000_1828140.flac	30,548	2,073,683	14.14	true	true	true	true	null	null
4405_2073683_1830480_1835060.flac	4,405	2,073,683	4.58	false	false	true	true	null	null
30065_2073683_1835080_1861940.flac	30,065	2,073,683	26.86	true	true	true	true	null	null
30065_2073683_1861940_1885940.flac	30,065	2,073,683	24	true	true	true	true	null	null
30065_2073683_1885940_1900040.flac	30,065	2,073,683	14.1	true	true	true	true	null	null
4405_2073683_1917060_1919340.flac	4,405	2,073,683	2.28	false	false	true	true	null	null
30065_2073683_1921200_1934240.flac	30,065	2,073,683	13.04	true	true	true	true	null	null
30065_2073683_1954150_1972920.flac	30,065	2,073,683	18.77	true	true	true	true	null	null
4405_2073683_1973880_1974960.flac	4,405	2,073,683	1.08	false	false	true	true	null	null
30058_2073683_1975880_1995280.flac	30,058	2,073,683	19.4	false	false	true	true	null	null
30058_2073683_1995280_2019760.flac	30,058	2,073,683	24.48	true	true	true	true	null	null
4405_2073683_2022080_2023300.flac	4,405	2,073,683	1.22	false	false	true	true	null	null
30058_2073683_2025600_2046180.flac	30,058	2,073,683	20.58	false	false	true	true	null	null
4405_2073683_2046880_2055560.flac	4,405	2,073,683	8.68	true	true	true	true	null	null
30058_2073683_2056100_2069520.flac	30,058	2,073,683	13.42	false	false	true	true	null	null
30058_2073683_2082820_2112800.flac	30,058	2,073,683	29.98	false	false	true	true	null	null
30058_2073683_2112800_2139640.flac	30,058	2,073,683	26.84	true	true	true	true	null	null
30058_2073683_2139640_2162620.flac	30,058	2,073,683	22.98	true	true	true	true	null	null
4405_2073683_2163940_2182200.flac	4,405	2,073,683	18.26	false	false	true	true	null	null
30095_2073683_2183180_2212100.flac	30,095	2,073,683	28.92	true	true	true	true	null	null
30095_2073683_2212100_2241080.flac	30,095	2,073,683	28.98	true	true	true	true	null	null
30095_2073683_2241080_2269980.flac	30,095	2,073,683	28.9	true	true	true	true	null	null
30095_2073683_2269980_2296340.flac	30,095	2,073,683	26.36	true	true	true	true	null	null
30095_2073683_2296340_2303240.flac	30,095	2,073,683	6.9	true	true	true	true	null	null
4405_2073683_2303240_2312420.flac	4,405	2,073,683	9.18	false	false	true	true	null	null
2178_2073683_2312560_2334500.flac	2,178	2,073,683	21.94	false	false	true	true	null	null
2178_2073683_2334500_2360560.flac	2,178	2,073,683	26.06	false	false	true	true	null	null
2178_2073683_2360560_2381920.flac	2,178	2,073,683	21.36	false	false	true	true	null	null
2178_2073683_2381920_2393920.flac	2,178	2,073,683	12	false	false	true	true	null	null
4405_2073683_2440340_2444240.flac	4,405	2,073,683	3.9	true	true	true	true	null	null
2178_2073683_2444340_2445580.flac	2,178	2,073,683	1.24	false	false	true	true	null	null
4405_2073683_2446260_2448000.flac	4,405	2,073,683	1.74	false	false	true	true	null	null
2178_2073683_2449980_2475400.flac	2,178	2,073,683	25.42	false	false	true	true	null	null
4405_2073683_2509780_2511840.flac	4,405	2,073,683	2.06	false	false	true	true	null	null
2178_2073683_2512240_2539600.flac	2,178	2,073,683	27.36	false	false	true	true	null	null
2178_2073683_2539600_2567640.flac	2,178	2,073,683	28.04	false	false	true	true	null	null
2178_2073683_2567640_2594060.flac	2,178	2,073,683	26.42	false	false	true	true	null	null
2178_2073683_2594060_2622840.flac	2,178	2,073,683	28.78	true	true	true	true	null	null
2178_2073683_2622840_2652300.flac	2,178	2,073,683	29.46	true	true	true	true	null	null
4405_2073683_2653240_2655300.flac	4,405	2,073,683	2.06	true	true	true	true	null	null
2178_2073683_2655480_2684060.flac	2,178	2,073,683	28.58	false	false	true	true	null	null
4405_2073683_2684540_2688960.flac	4,405	2,073,683	4.42	false	false	true	true	null	null
2178_2073683_2689440_2712660.flac	2,178	2,073,683	23.22	false	false	true	true	null	null
2178_2073683_2712660_2728660.flac	2,178	2,073,683	16	false	false	true	true	null	null
4405_2073683_2728920_2747740.flac	4,405	2,073,683	18.82	false	false	true	true	null	null
4405_2073683_2787140_2788900.flac	4,405	2,073,683	1.76	false	false	true	true	null	null
4405_2073683_2844660_2847600.flac	4,405	2,073,683	2.94	false	false	true	true	null	null
4405_2073683_2857960_2861080.flac	4,405	2,073,683	3.12	false	false	true	true	null	null
4405_2073683_2868740_2874380.flac	4,405	2,073,683	5.64	false	false	true	true	null	null
30095_2073683_2874880_2883560.flac	30,095	2,073,683	8.68	true	true	true	true	null	null
2178_2073683_2884940_2911080.flac	2,178	2,073,683	26.14	true	true	true	true	null	null
2178_2073683_2911080_2926620.flac	2,178	2,073,683	15.54	false	false	true	true	null	null
4405_2073683_2927960_2934420.flac	4,405	2,073,683	6.46	false	false	true	true	null	null
4405_2073683_3020280_3024820.flac	4,405	2,073,683	4.54	true	true	true	true	null	null
4405_2073683_3069480_3094460.flac	4,405	2,073,683	24.98	false	false	true	true	null	null
2178_2073683_3095640_3123960.flac	2,178	2,073,683	28.32	false	false	true	true	null	null
4405_2073683_3124560_3136960.flac	4,405	2,073,683	12.4	false	false	true	true	null	null
2178_2073683_3138040_3139780.flac	2,178	2,073,683	1.74	false	false	true	true	null	null
4405_2073683_3324300_3325920.flac	4,405	2,073,683	1.62	false	false	true	true	null	null
4405_2073683_3345540_3353860.flac	4,405	2,073,683	8.32	false	false	true	true	null	null

End of preview. Expand in Data Studio

Knesset Committees Inference

Transcriptions of Knesset committee audio by two models, with the protocol reference alongside, for the personalised-ASR study (Stage 1: per-speaker WER, general model vs Hebrew fine-tune). Audio and references come from Hadasy/knesset-committees-chunks; speaker identities from Dolevabudi/knesset-committees-speakers. No audio is included.

arm	model	served by	language
A	`openai/whisper-large-v3`	HF Inference (deepinfra)	forced `he`
B	`ivrit-ai/whisper-large-v3-turbo-ct2`	RunPod serverless (faster-whisper)	forced `he`
A_auto	`openai/whisper-large-v3`	same as A	auto-detected (see below)

Files

inference.parquet -- one row per chunk, hypothesis_A and hypothesis_B side by side with reference, speaker_id, session, session_date, knesset, duration_s, quality; in_stage1_subset marks the evaluated subset. 65,990 subset chunks, 230 h, 267 speakers, all with both hypotheses.
inference_long.parquet -- one row per (chunk, arm) with timing (latency_s, provider exec_s/queue_s) and any error; 219,180 rows.
coverage.parquet -- one row per chunk of the whole corpus (1,204,617): in_subset, done_A, done_B, done_A_auto, errors. The record of what was and was not transcribed.

The subset

One hour per MK (alignment quality >= 0.5), drawn round-robin over sessions so it spans the speaker's tenure. Four speakers have under six minutes of usable audio in total and correspondingly few chunks (speaker_id 12944, 30684, 30698, 30846); filter by hours before per-speaker statistics.

Language forcing and `hypothesis_A_auto`

The first Arm A run let Whisper detect the language. On chunks under 3 s it guessed wrong a quarter of the time (Portuguese, Arabic, Russian...), 6.7 % of all chunks. Arm B is always told the language, so the comparison was unfair; Arm A was re-run with language='he'. The re-run is hypothesis_A (arm A); the auto-detect run is kept as hypothesis_A_auto (arm A_auto in the long table) because the detection-failure rate is itself a finding.

Scoring as validated

normalize_he (stage2/pipeline.py) then word-level Levenshtein, corpus WER on the subset: A 0.417, B 0.324. Empty hypotheses: A 0.30 %, B 0. Hebrew script: A 99.7 %, B 99.99 %.

Licence

References and speaker attribution derive from the Knesset Corpus (CC-BY-SA 4.0); audio from ivrit.ai. This dataset is released under CC-BY-SA 4.0.

Downloads last month: 19