UDGAM LABS · Dataset · what 27 speech-to-text settings write on audio with no one speaking

Hush-Bench

A good speech-to-text model writes nothing when nobody speaks. Hush-Bench measures how often models write words anyway. It has 1,000 audio clips, every one CC0, public domain or generated by us. 950 of them (5.1 hours) are scored:

Set	Scored clips	What it holds
Silence	150	digital and near-digital silence, room tone, mains hum, clicks, codec and phone-band silence, dropouts, 10 long silences of 1 to 5 minutes
Noise	296	vehicles, engines, clocks, bells, streams, rain, printers and other everyday sounds; generated coloured, modulated and machine noise and crackle
Music	213	instrumental recordings (orchestra, piano, band, guitar, electronic and more) and short synthetic pieces
Non-verbal	91	laughter, throat clearing, sneezes, sighs, gasps, snoring, a baby's cry, clapping, footsteps: human sounds without words
Speech + silence	200	one spoken English sentence before, after or between stretches of the above

The headline set is the 659 clips with no human voice at all (silence, noise, music). Non-verbal sounds are reported separately, because writing "ha ha" for a laugh is arguably a transcript. The speech set checks that a model which writes nothing on silence still transcribes speech.

No person listened to the clips. The other 50 clips (34 non-verbal, 12 music, 4 noise) were flagged by the automatic screen or by the models' own outputs as possibly holding a voice. They are set aside and unverified: the audio is here, they are listed in holdout.csv, and no number below counts them.

Quick start

import csv
import soundfile as sf
from huggingface_hub import snapshot_download

root = snapshot_download("udgamlabs/hush-bench", repo_type="dataset")
set_aside = {r["id"] for r in csv.DictReader(open(f"{root}/holdout.csv", encoding="utf-8"))}
for row in csv.DictReader(open(f"{root}/clips.csv", encoding="utf-8")):
    if row["id"] in set_aside:
        continue                                        # flagged, unverified, not scored
    audio, sr = sf.read(f"{root}/{row['file']}")       # mono, 16 kHz
    text = my_transcribe(audio, sr)                     # your system here
    print(row["id"], row["category"], repr(text))

A clip "has invented words" if the transcript, after removing caption tags such as [Music], (applause), *Loud noise* and music-note symbols, still contains a word. Chinese and Japanese characters count one word each. scoring/ holds the exact scorer used below.

Results

Each model at a pinned revision with its model card's recommended decoding, run on 23–24 October 2026 and scored on the 950 clips on 8 November 2026. Intervals are 95% (Wilson for shares, bootstrap for words per minute). Speech test: word error rate on the 200 speech-plus-silence clips. Every transcript is in outputs/; every table in results/.

System	No-voice clips with invented words	Words per minute	Silence	Noise	Music	Non-verbal (91 scored)	Non-verbal, set-aside included (125)	Speech-test WER
Kyutai STT 1B (en/fr)	0/659 = 0.0% (0.0–0.6%)	0.0	0%	0%	0%	0%	0.8%	26.3%
Nemotron 3.5 ASR streaming 0.6B	0/659 = 0.0% (0.0–0.6%)	0.0	0%	0%	0%	2.2%	2.4%	16.9%
Qwen3-ASR 1.7B	0/659 = 0.0% (0.0–0.6%)	0.0	0%	0%	0%	7.7%	12%	8.1%
Voxtral Mini 4B Realtime	0/659 = 0.0% (0.0–0.6%)	0.0	0%	0%	0%	0%	0%	13.0%
Canary 1B v2	1/659 = 0.2% (0.0–0.9%)	0.1	0%	0.3%	0%	0%	3.2%	9.0%
Moonshine Streaming tiny	3/659 = 0.5% (0.2–1.3%)	0.0	2.0%	0%	0%	0%	0%	34.7%
Parakeet TDT 0.6B v3	8/659 = 1.2% (0.6–2.4%)	0.1	1.3%	1.0%	1.4%	2.2%	6.4%	8.7%
Qwen3-ASR 0.6B	10/659 = 1.5% (0.8–2.8%)	0.0	3.3%	0%	2.3%	11%	20%	9.9%
Granite Speech 5.0 470M CTC	15/659 = 2.3% (1.4–3.7%)	0.1	8.0%	1.0%	0%	7.7%	18%	6.9%
Moonshine Streaming small	23/659 = 3.5% (2.3–5.2%)	17.7	8.7%	1.0%	3.3%	0%	0.8%	22.2%
Moonshine Streaming medium	27/659 = 4.1% (2.8–5.9%)	0.9	0%	0.3%	12%	0%	0%	19.9%
Whisper large-v3-turbo, faster-whisper with its own VAD	47/659 = 7.1% (5.4–9.4%)	0.4	0%	2.7%	18%	23%	34%	11.6%
Shruti (Udgam Labs)	111/659 = 17% (14–20%)	1.7	0%	23%	20%	53%	60%	31.0%
Vosk small en	164/659 = 25% (22–28%)	1.4	29%	22%	25%	26%	35%	37.9%
wav2vec2-base-960h	185/659 = 28% (25–32%)	5.7	0%	26%	50%	48%	54%	61.1%
Whisper large-v3-turbo, English set	381/659 = 58% (54–62%)	10.5	47%	59%	64%	71%	76%	13.0%
Whisper large-v3-turbo (transformers)	384/659 = 58% (54–62%)	13.1	47%	59%	66%	74%	78%	13.1%
Whisper large-v3	469/659 = 71% (68–74%)	47.4	79%	78%	56%	80%	85%	24.3%
Cactus Whistle	478/659 = 73% (69–76%)	5.8	19%	85%	93%	100%	100%	25.3%
Whisper large-v3-turbo, faster-whisper	550/659 = 83% (80–86%)	11.7	85%	80%	87%	91%	94%	18.9%
Whisper large-v3-turbo, openai-whisper	571/659 = 87% (84–89%)	14.2	76%	87%	94%	100%	99%	18.5%
Whisper medium	591/659 = 90% (87–92%)	165.7	65%	96%	99%	97%	97%	27.5%
Whisper tiny	608/659 = 92% (90–94%)	170.3	97%	89%	93%	85%	88%	47.4%
Whisper base	622/659 = 94% (92–96%)	182.0	99%	98%	86%	93%	95%	39.0%
Whisper small	629/659 = 95% (94–97%)	31.4	91%	95%	100%	99%	99%	28.8%
Distil-Whisper large-v3.5	652/659 = 99% (98–99%)	5.1	100%	98%	100%	98%	98%	9.8%
Granite Speech 5.0 2B	654/659 = 99% (98–100%)	140.4	99%	99%	100%	98%	98%	38.4%

How the 50 were chosen, and what it does to the numbers. All 16 set-aside no-voice clips (12 music, 4 noise) and 17 of the 34 set-aside non-verbal clips were flagged because models wrote words on them. Setting them aside therefore favours the models that wrote on them. On the no-voice clips the effect is small: no share moves by 1.3 points or more, and Qwen3-ASR 1.7B's 3 clips are the reason it now shows zero. On non-verbal sounds it is large for the careful models, so the table shows both ways: Parakeet TDT v3 6.4% with the set-aside clips and 2.2% without, Granite Speech 470M CTC 18% and 7.7%, Shruti 60% and 53%. Every system's output on the 50 is in results/per_clip_held_out.csv.

• Three models wrote nothing on any no-voice clip, the set-aside ones included: Kyutai STT 1B, Nemotron 3.5 ASR and Voxtral Mini 4B Realtime. Qwen3-ASR 1.7B wrote nothing on the 659 scored no-voice clips, and short words ("the", "oh oh oh") on 3 set-aside music clips.
• Whisper invents text on most no-voice clips in every size and runtime we ran. The commonest invention across all models is "Thank you." (on 684 of the 750 scored non-speech clips, by at least one model), then "you", "Music", "The End", "I'll see you next time", "Thank you for watching" and the subtitle credit "Sous-titrage ST' 501". Smaller Whisper models fall into repetition loops of 165–182 words per minute.
• Granite Speech 5.0 2B wrote whole sentences on silence ("The company's revenue for the quarter was 1.2 billion yen", on 10 clips) and 9,664 words on the 11 long no-voice clips.
• The runtime matters. The same Whisper large-v3-turbo weights invent text on 58%, 83% or 87% of no-voice clips depending on the library, and on 7.1% with faster-whisper's built-in VAD.

With a voice-activity detector in front

Silero VAD v6.2 (threshold 0.5) keeps only the stretches it judges to be speech; a clip with none is not sent to the model. With it, every system invents text on 1.8% or fewer of the no-voice clips: the VAD passes 12 of the 659 (11 music recordings and a foghorn, each with under 0.5 s judged as speech). On the speech test it repairs the models that loop (Granite Speech 2B: 38.4% to 8.1% word error rate; faster-whisper turbo: 18.9% to 11.0%) and costs the careful ones a little (Parakeet 8.7% to 10.2%, Qwen3-ASR 1.7B 8.1% to 10.1%). Every system with and without the VAD, with paired tests, is in results/tables_full.md.

Where it falls short

• One setting per model, one revision, one GPU type, plus four runtimes of Whisper large-v3-turbo to show how much the runtime matters. Other decoding or language settings can give different numbers. Language was detected automatically unless the model card sets it.
• Our clips, not your audio. Phone calls, meetings, TV and background chatter are not covered, and only 11 scored no-voice clips are longer than a minute.
• No person listened to the clips. Every recording passed an automatic speech and singing screen, and the 50 clips flagged by the screen or the models' outputs are set aside. A faint voice that got through would count against every model alike. Four clips the models flagged stay in the scored set because they were in a random sample kept for checking (mus-r072, mus-r084, nvb-r008, nvb-r094); five speech-test clips use a set-aside clip as their background sound; four scored clips are cut from the same recordings as set-aside clips.
• Not every open model is here. The models ran on 23–24 October 2026. Models released since, and some released shortly before, are not on the board.
• The speech test is English (Mozilla Common Voice). Some models list Common Voice in their training data.
• Counting rules matter. Caption tags are not counted as words; results/ also has the strict view that counts them.
• Shruti is our own model (built on Cactus Whistle). It ran under the same rules as every other system.

How the clips were made

These are the full 1,000, before the 50 are set aside.

• 235 generated by us and released as CC0: silence and near-silence of many kinds, hum, clicks, dropouts, coloured and modulated noise, machine whine, crackle, and 25 short synthetic instrumentals. clips.csv has the recipe.
• 401 cuts from 326 Wikimedia Commons recordings that are CC0 or in the public domain, with the licence read from each file's Commons page and the source file's SHA-1 checked.
• 164 FSD50K clips whose own Freesound licence is CC0 and whose labels contain no speech, singing or crowd class.
• 200 speech clips combine one sentence from the Mozilla Common Voice 17.0 English test set (CC0) with the material above; clips.csv gives the reference text and where the speech starts and ends.
• Every recording was screened by an AudioSet tagger and a voice-activity detector, with thresholds calibrated on known speech and singing.
• On 8 November 2026 we re-read the licence of a random 40 recordings at their source (25 Commons, 15 Freesound): all 40 match clips.csv (results/licence_spotcheck.json).
• Every clip is mono, 16 kHz, 16-bit WAV. clips.csv lists each clip's licence, source page, author and the SHA-256 of our file; vad.csv the VAD's speech segments.

Files

• clips/ and clips.csv: all 1,000 clips and their provenance. vad.csv: Silero VAD segments per clip.
• holdout.csv: the 50 set-aside clips and why each was flagged. results/per_clip_held_out.csv: what every system wrote on them, kept apart from every score.
• outputs/: every system's raw transcript of all 1,000 clips (<system>.jsonl) and its exact model revision, file hashes, library versions and decoding settings (<system>.meta.json).
• results/: the headline table, every category, every pair of systems, the VAD effect, phrase counts and per-clip scores.
• scoring/: the scorer (score_full.py, headline.py, with the shared helpers).

Related work

Other public work looks at the same failure from other sides. HALAS has people mark invented text in seven models' transcripts of real earnings calls, so it covers invention on top of real speech, which Hush-Bench does not. How to Reduce Whisper Hallucination releases a larger benchmark of 11,852 clips built to compare Whisper fine-tunes. Hush-Bench puts many different open models and runtimes on one smaller clip set, with and without a VAD, and publishes every transcript.

Licence

The clips we generated are dedicated to the public domain under CC0 1.0. The recordings are CC0 or public domain as stated at their source (clips.csv links each one); Common Voice sentences are CC0. We add no rights of our own. Public-domain status can differ by country, so check clips.csv if that matters for your use. Authors are credited in clips.csv even where no credit is required.

Citation

@misc{udgamlabs2026hushbench,
  title  = {Hush-Bench: words invented by speech-to-text models on audio with no speech},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/udgamlabs/hush-bench}
}

On udgamlabs.com: Blog post

---

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com