Scicom-intl/Whisper-Hallucination

Whisper Hallucination and Repetition Probes

This is a BENCHMARK. Every evaluation config is test — do not fine-tune on it. (The one exception is lexicon_synth, which is synthetic training material and ships its own train/test split. It is not one of the eight benchmark arms — see below.) Training on these clips invalidates every number you would then report. Build training data separately from the same source corpora, excluding the items listed in benchmark/exclusions.json in the project repo (546 FMA tracks, 1,168 FSD50K ids, 2,620 LibriSpeech utterances, and the Malay/Lingua Libre stems).

Structural disjointness is easy here: the benchmark draws FSD50K's eval split and FMA shards 0-1, so training can use FSD50K dev (35,676 voice-free clips) and FMA shards 2-12 and never overlap. Malaysian-Emilia (3,727 h, 17,290 genuine terima kasih) is untouched by this benchmark and is the natural positive pool.

Audio probes and phrase lexicons for measuring the two failure modes of whisper-large-v3:

• hallucination — text with no phonetic basis in the audio (Terima kasih. over silence)
• repetition / looping — the decoder emits a unit far more times than it was spoken

Code, build scripts and design notes: github.com/Scicom-AI-Enterprise-Organization/Whisper-Hallucination — arm builders, the benchmark harness, benchmark/exclusions.json, the ablation grid and the training-corpus tooling all live there.

The problem this is built around

Every high-frequency Whisper hallucination is also a phrase people genuinely say. terima kasih is the most common Malay hallucination and the most common thing a Malaysian call-centre agent says. You cannot fix this with a text blocklist without deleting real transcriptions.

And silence is not the only trigger. Music and noise hallucinate at least as readily — Calm-Whisper measured large-v3 producing text on 99.97% of UrbanSound8K clips, none of which contain speech. Energy in the signal is not evidence for words.

So the dataset is contrastive, across a spectrum rather than a binary:

state	audio	Whisper says	correct action
silence	room tone, no speech	Terima kasih.	drop — 100% invented
music / noise	a track, no voice	Terima kasih.	drop — energy, no evidence
speech in noise	real speech at low SNR	part real, part invented	the hard case
clean speech	someone says it	Terima kasih.	keep — correct

A detector that only reads the text cannot tell these apart. That is the point.

Baseline scores

Five checkpoints × 8 arms × 11,852 clips per model, measured 2026-09-16/17 on 2× NVIDIA H20 with greedy decoding and no temperature fallback: large-v2, large-v3, large-v3-turbo, and the two Malaysian fine-tunes mesolitica/malaysian-whisper-large-v2 and mesolitica/Malaysian-whisper-large-v3-turbo-v3.

Full tables — hallucination rates, repetition runaway, WER, and what each model actually emits on silence and music — are in the project repo: README → Measured baselines. Raw numbers: bench/scores.json.

Headline: on non-speech, hallucination runs 9.3-98.8% depending on model and arm. For the three OpenAI checkpoints it is 100.0% everywhere once a bare "." counts as output — every clip produces something. The Malaysian fine-tunes break that: Malaysian-turbo-v3 returns a genuinely empty string on 90.6% of voice-free clips. It pays for that on the other axis, running away on 14.5% of reduplication clips against 5.8% for the checkpoint it came from, and emitting nothing on 58% of them. LibriSpeech test-clean sits at 3.5% for large-v3, matching the published figure — the check that the harness is wired correctly.

Measuring both failure modes on the same clips is the point: ranked on hallucination alone Malaysian-turbo-v3 wins by a distance; ranked on raw WER it looks unusable, and ~1% of looping clips account for most of that gap.

The benchmark harness (bench/) ships with this dataset and in the repo.

Configs

config	rows	hours	content	ground truth
reduplication	1,440	1.9	repeated units — CV syllables, vowels, laughter, clicks	exactly n_repeats repeats
silence	42	0.4	silence / near-silence, 6 floors × 7 durations	empty string
music	600	4.5	real produced music — Free Music Archive excerpts	empty string
nonspeech	1,168	2.8	FSD50K music + environmental noise, label-verified voice-free	empty string
speech_in_noise	1,200	3.1	genuine speech mixed with music/noise at 5 SNRs	the real transcript
genuine	4,694	12.4	real Malaysian speech, permissively licensed	human transcript
genuine_isolated	88	0.05	native speakers saying one phrase, nothing else	the phrase
librispeech_test_clean	2,620	5.4	English WER regression guard	human transcript
lexicon	40,891	—	known hallucination phrases, 100 languages	—
ban_candidates	40,891	—	every lexicon phrase, classified safe/unsafe to blocklist	—
targets	463	—	high-risk phrases: hallucinated and genuinely said	—
malaysian_sources	24	—	survey of public Malaysian speech datasets	—

from datasets import load_dataset

# Audio arms are test-only.
red = load_dataset("Scicom-intl/Whisper-Hallucination", "reduplication",   split="test")
sil = load_dataset("Scicom-intl/Whisper-Hallucination", "silence",         split="test")
mus = load_dataset("Scicom-intl/Whisper-Hallucination", "music",           split="test")
sin = load_dataset("Scicom-intl/Whisper-Hallucination", "speech_in_noise", split="test")
gen = load_dataset("Scicom-intl/Whisper-Hallucination", "genuine",         split="test")

# Lookup tables (not splits).
lex = load_dataset("Scicom-intl/Whisper-Hallucination", "lexicon",         split="train")
ban = load_dataset("Scicom-intl/Whisper-Hallucination", "ban_candidates",  split="train")

Decoding the audio

datasets >= 5.0 raises ImportError: To support decoding audio data, please install 'torchcodec' the moment you touch an audio column. Either install it, or skip it entirely — the payload is 16 kHz mono FLAC that soundfile reads directly:

import io, soundfile as sf
from datasets import load_dataset, Audio

ds = load_dataset("Scicom-intl/Whisper-Hallucination", "reduplication", split="test")
ds = ds.cast_column("audio", Audio(decode=False))          # hand back raw bytes

row = ds[0]
x, sr = sf.read(io.BytesIO(row["audio"]["bytes"]), dtype="float32")   # 16 kHz mono
print(row["id"], row["n_repeats"], row["reference_text"], len(x) / sr)

streaming=True works the same way, if you would rather not pull 2 GB to look at a few clips:

st = load_dataset("Scicom-intl/Whisper-Hallucination", "silence", split="test", streaming=True)
st = st.cast_column("audio", Audio(decode=False))
row = next(iter(st))

Audio is 16 kHz mono FLAC (lossless, 16-bit), peak-normalised to −3 dBFS.

lexicon_synth — synthetic positives, with a train split

Not a benchmark arm. The eight audio configs above are evaluation-only. This one is material for building a mitigation: 16,922 synthesised clips of lexicon phrases, 8.6 hours, across 83 languages, shipped as train (13,203 clips) and test (3,719).

	train	test
clips	13,203	3,719
hours	6.74	1.84
distinct phrases	10,413	2,883
languages	75	83
languages with ≥10 clips	72	70
non-English share	86%	85%

The split is by PHRASE, not by clip. Each phrase is rendered in several voices; splitting clips would put terima kasih in both halves and a model would then be evaluated on text it trained on. Assignment is a hash of (phrase, language) with a fixed salt, so it is reproducible without shipping a map and stable as the corpus grows.

306 phrases are forced into test because they appear in phrases/targets.csv, which drives the genuine_isolated arm. Synthetic audio of a benchmark phrase contaminates the benchmark exactly as its real audio would, so those phrases can never enter train. The build re-verifies both properties — zero phrase overlap, zero benchmark phrases in train — and refuses to write a release if either fails.

How it was made. Phrases come from three places, and meta.phrase_provenance says which for every clip:

provenance	count	what it is
observed	upstream lexicons	phrases Whisper was seen to hallucinate
mined	our own runs	what these five checkpoints emitted on our non-speech arms
translated	NLLB-200	English hallucination phrases rendered into 80 languages

translated rows are positives, not evidence: a Tamil thank you is something people say, not something Whisper was observed to invent in Tamil. Do not feed them to ban_candidates — that would be circular.

Audio is from Scicom-intl/Multilingual-Expressive-TTS-1.7B (speaker-name conditioned, 45 verified voices) and k2-fsa/OmniVoice, routed per language by measured CER. Every clip's meta column is a JSON string carrying the TTS model, codec, conditioning mode, speaker, sample rates, decoding parameters and seed.

Every clip passed an ASR round-trip filter — Whisper must recover the phrase, and the clip must be about as long as the phrase should take, which rejects a TTS that keeps talking past a two-word prompt. The gate is per language, anchored to the judge's own floor: we measured Whisper's CER on real FLEURS speech in 44 languages, and a language where the judge itself scores 0.88 (Burmese) or 1.23 (Amharic) cannot be held to a 0.25 gate. Nine such languages are excluded rather than scored, because their clips are unverifiable with this judge, not necessarily bad.

train = load_dataset("Scicom-intl/Whisper-Hallucination", "lexicon_synth", split="train")
test  = load_dataset("Scicom-intl/Whisper-Hallucination", "lexicon_synth", split="test")

Negative arms — what the model invents

reduplication

Motivated by an observed failure: a recording of the syllable tu repeated four times decoded as tu repeated until the token limit. Once inside a repeated-token region, Whisper's decoder has no signal for how many repeats remain.

field	values
pattern	cv (720), vowel (360), laugh (216), click (144)
unit	tu ta ka pa da la na bi ko me / a i u e o / ha he hi / tick beep
n_repeats	3, 4, 5, 6, 8, 12
rate_hz	3, 5, 7, 9 units/sec
tail_silence_s	0, 2, 8

click contains no speech at all, which separates repetition from speech as the trigger. Nothing in this config is tied to a language.

Score with n_hyp / n_repeats: 1.0 correct, >1 a runaway, <1 a mitigation that deleted genuine repetition.

silence

The cleanest arm to read: the reference is the empty string, so any output is a hallucination. floor covers digital_zero, dither, hiss_-60db, hiss_-45db, hum_50hz, roomtone — true digital zero behaves differently from a realistic quiet mic. Durations straddle the 30 s window boundary (1, 5, 10, 29, 31, 60, 120 s).

music — 600 excerpts, 4.5 h

Real produced music, not sound effects: Free Music Archive tracks (commercial-licence subset, CC BY 4.0, already 16 kHz mono), excerpted at 10 / 30 / 45 s from inside the track so intros are skipped. This is what actually plays under a call — hold music, a YouTube backing bed — and it is a different acoustic distribution from isolated instrument samples.

Reference is the empty string: music is not speech, so a transcript is a hallucination.

Vocals caveat. FMA ships no instrumental/vocal tag, so some tracks contain singing. The vocals column records unknown rather than a guess. Those rows still carry an empty reference, because the target behaviour for a music bed is to emit nothing — but if you need provably zero words in the audio, use nonspeech, where every clip is label-verified voice-free.

nonspeech — 1,168 clips, 2.8 h

FSD50K clips (CC BY 4.0) carrying no voice label of any kind — 583 music / musical-instrument, 585 environmental and mechanical noise. Any clip labelled Speech, Singing, Human_voice, Chatter, Laughter and 20-odd relatives was excluded, so the empty reference is verified by annotation, not assumed.

This is the strict-ground-truth non-speech arm; music is the realistic one.

speech_in_noise — 1,200 clips, 3.1 h

The case the pure arms cannot reach: genuine speech with a background bed, where transcription degrades into fabrication as evidence weakens rather than appearing from nothing. 240 clips from genuine (preferring those carrying a high-risk phrase) mixed with voice-free FSD50K music or noise at SNR = +20, +10, +5, 0, −5 dB.

Reference is the real transcript, so this measures WER degradation and hallucination onset on the same axis. 295 of the 1,200 carry a target phrase — those are where a hallucinated terima kasih and a real one become genuinely confusable.

Positive arms — what a mitigation must not destroy

Without these, "hallucination rate fell" is unfalsifiable: a decoder that outputs nothing scores perfectly on the negative arms.

genuine — 4,694 clips, 12.4 h

source	clips	licence	register
SaLTUNIMAS/sarawak-malay-asr	1,164	CC BY 4.0	Sarawak Malay interviews, human transcripts
emhaihsan/Synth-Manglish	2,457	CC BY 4.0	Manglish code-switching (synthetic TTS voice)
google/fleurs ms_my	1,073	CC BY 4.0	human-read Wikipedia prose

59 clips carry a high-risk phrase (has_target_phrase = true) — these are the real minimal pairs against the negative arms:

phrase	clips
sama-sama	41
terima kasih	16
selamat tinggal	2
terima kasih banyak banyak	1

The rest are the regression guard: ordinary speech whose WER must not move when a mitigation is applied.

librispeech_test_clean — 2,620 clips, 5.4 h

LibriSpeech test-clean (CC BY 4.0), the split the hallucination-mitigation literature reports its cost on. Included so a WER figure measured here lands on the same axis as published results rather than only on our Malaysian arms.

Reference point — hallucination-space projection on large-v3 (arXiv:2609.04561):

variant	ESC-50 HR	LibriSpeech WER clean / other
baseline	44.25%	4.06% / 5.87%
gated projection	8.38%	6.17% / 6.57%
always-on projection	1.50%	12.95% / 13.13%

Suppressing hallucination is not free: always-on nearly eliminates it and triples clean WER. That exchange rate is the reason the positive arms exist.

genuine_isolated — 88 clips

Native speakers from Lingua Libre (via Wikimedia Commons) saying one phrase and nothing else, across ms / id / en / zh / ta — 3 bare terima kasih, plus thanks, okay, please, sorry, bye, 好, sama-sama, maaf, selamat, baik, tidak, ya. Small, but the closest possible match to the hallucination's acoustic shape: a hallucination surfaces as a bare Terima kasih., and so does one of these clips. In genuine the phrase sits mid-sentence, which is the easier case.

Licences vary per speaker (CC0 ×43, CC BY-SA 4.0 ×42, CC BY 4.0 ×3), so every row carries its own license, author and source_url. Filter license != "CC BY-SA 4.0" if you need to avoid copyleft.

ban_candidates — which phrases you can actually blocklist

The founding problem, answered with measurement instead of intuition. Every lexicon phrase is crossed against all 7,402 genuine transcripts here (17.8 h) and classified:

verdict	phrases	meaning
safe_candidate	871	never observed in genuine speech, multi-word, frequently hallucinated
unsafe	389	occurs in genuine speech — banning it destroys real transcripts
weak_evidence	39,631	never observed, but too rare or too short to be confident

phrase	verdict	hallucinated	genuine exact / substring
thanks for watching	safe_candidate	25,053	0 / 0
thank you	unsafe	31,353	0 / 13
terima kasih	unsafe	1	3 / 19

Use it in EXACT mode — drop an output only when the whole transcript equals the phrase, which is the shape a hallucination takes. genuine_substring shows what a more aggressive contains-match would additionally cost.

Caveat, and it is the important one: every ms / zh / ta phrase falls into weak_evidence, because the hallucination counts come from public lexicons that barely cover those languages (6 Malay phrases in total). The classifier is sound — English demonstrates it — but the Malaysian side needs hallucination counts measured locally, which this dataset's probe arms exist to produce.

lexicon

Four public sources merged, normalised (NFC, casefold, punctuation-stripped) and deduplicated on (phrase, lang):

source key	rows in	origin	licence
agh_boh	294	AGH DSP "Bag of Hallucinations", ICASSP 2025	MIT
agh_full	30,407	same paper, full non-speech tally	MIT
hf_noise	7,889	sachaarbonel/whisper-hallucinations	MIT
granary	3,028 / 23 langs	NVIDIA NeMo SDP, Granary pipeline	Apache-2.0

Top English entries: thank you (31,353), thanks for watching (25,053), thank you for watching (6,264). The NVIDIA lists ship alongside a working DetectWhisperHallucinationFeatures processor, vendored at lexicon_raw/granary/.

Coverage is extremely uneven. English has 30,443 phrases; Malay has 6. That gap is why this dataset exists — the Malaysian side is not downloadable, it has to be measured.

A note on comparing hallucination rates

Published HRs for the same model on the same corpus differ by more than 10x — Calm-Whisper reports 99.97% on UrbanSound8K for large-v3, arXiv:2609.04561 reports 76.08%. The gap is methodological: the latter counts only output surviving Whisper's own no-speech filter (no_speech_threshold = 0.6), the former counts raw output. Always state which you mean.

It also matters what is in the corpus. arXiv:2609.04561 filtered FSD50K down to clips labelled neither speech, vocal nor music (HR 21.35%). Our nonspeech arm deliberately keeps music — filter kind == "noise" to approximate their setup, or use the whole arm plus music for the harder, more realistic condition.

Known limitations

• The positive arms are small relative to the negative ones, and the phrase-matched subset is 147 clips (59 + 88). That is what genuinely exists under a redistributable licence. The larger Malaysian conversational corpora are CC BY-NC and were excluded on purpose so this dataset stays shareable.
• Synth-Manglish is TTS, not human speech. It is a little over half of genuine. Filter on source if that matters for your claim.
• genuine clips are long (median 9.6 s) while hallucinations happen on short, low-evidence audio. The isolated arm is the short-clip counterpart.
• music may contain singing. See the vocals caveat above; nonspeech is the label-verified alternative.
• speech_in_noise is synthetic mixing, not natively recorded noisy speech. SNR is exact and controllable, which is the point, but channel effects and Lombard speech (people talk differently in noise) are absent.
• No ablation results here. See the design notes in the project repo: ABLATION.md, TRAINING.md.
• FLEURS is human-read Wikipedia prose, so it contributes almost no conversational closers — 3,740 utterances yielded one terima kasih. Register matters more than size.

Files beyond the configs

lexicon_raw/            vendored upstream lexicons, unmodified (MIT / Apache-2.0)
  granary/<lang>.txt    NVIDIA NeMo per-language phrase lists, 23 languages
external/               HALAS (CC BY 4.0) — 3,611 Earnings-22 clips with human
                        hallucination / looping spans across 9 ASR models
licenses/               upstream licence texts + NOTICE.md

HALAS per-model "hallucination or looping" counts out of 3,611. Note that large-v3-turbo hallucinates more than large-v3:

model	flagged
whisper-large-v2	1,581
parakeet-tdt-v2	1,217
canary-1b	1,213
canary-1b-flash	1,099
phi-4	1,096
whisper-large-v3-turbo	1,060
granite	1,002
whisper-large-v3	858
CrisperWhisper	772

Provenance and licensing

Mixed, all permitting redistribution with attribution. Per component:

component	licence	attribution
reduplication, silence audio	CC BY 4.0	this dataset (synthesised, no recorded human speech)
music	CC BY 4.0	Free Music Archive (commercial subset); per-row artist / source_url
nonspeech	CC BY 4.0	FSD50K (Fhrozen/FSD50k)
speech_in_noise	CC BY 4.0	derived: genuine speech × FSD50K backgrounds
librispeech_test_clean	CC BY 4.0	LibriSpeech (openslr/librispeech_asr)
genuine	CC BY 4.0	per-row author / source_url
genuine_isolated	CC0 / CC BY 4.0 / CC BY-SA 4.0, per row	per-row author / source_url; Lingua Libre via Wikimedia Commons
lexicon (merged), targets, malaysian_sources	CC BY 4.0	this dataset
lexicon_raw/agh_*	MIT	AGH Signal Processing Group
lexicon_raw/hf_whisper_hallucinations_phrases.csv	MIT	Sacha Arbonel
lexicon_raw/granary/*	Apache-2.0	NVIDIA, NeMo-speech-data-processor
external/halas_*	CC BY 4.0	HALAS authors (AGH DSP)

The synthetic audio contains no recorded human speech — source-filter synthesis and generated noise — so it carries no speaker-consent or PII obligations. The genuine and genuine_isolated arms are redistributed from public corpora under their own licences; see licenses/NOTICE.md.

Content warning

The lexicon records what an ASR model wrongly generates on audio containing no speech. Of 40,891 phrases, roughly 586 contain profanity and 583 violence-related terms (oh shit, i m not sure if i can get the gun); 8 are sexual; none contain URLs. These are model errors, not anything a person said, and the upstream sources carry the same warning. They are kept because removing them would misrepresent the failure mode.

Citation

Dataset: Scicom-intl/Whisper-Hallucination. Code: https://github.com/Scicom-AI-Enterprise-Organization/Whisper-Hallucination

Built on:

• Koenecke, Choi, Mei, Schellmann, Sloane. Careless Whisper: Speech-to-Text Hallucination Harms. FAccT 2024.
• Barański et al. Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio. ICASSP 2025. (arXiv:2501.11378)
• Wang et al. Calm-Whisper: Reduce Whisper Hallucination On Non-Speech By Calming Crazy Heads Down. Interspeech 2025. (arXiv:2505.12969)
• HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems.
• NVIDIA NeMo-speech-data-processor, Granary pipeline.
• Lingua Libre / Wikimedia Commons contributors.