Namecoach/netjets-name-asr

NetJets Spoken-Name ASR + Phonetic Dataset

A dataset of spoken personal names (call-center / voicemail audio) paired with a gold name string and a gold phonetic transcription. It is intended for grapheme-to-phoneme (G2P), name-pronunciation, and robust short-utterance ASR research.

Splits

Split	Rows	Description
train	36,704	Clips with a usable transcription. name is reconstructed from what was actually spoken.
empty	1,120	Clips where the ASR produced no usable output (silence / heavily degraded). name falls back to the reference name. Useful as a hard-negative / evaluation set.

An additional 151 source recordings were missing from storage and are not included.

Features (in order)

Field	Type	Description
audio	audio	16 kHz mono WAV, one spoken utterance (typically ~2–4 s). Playable below.
Gemini IPA	string	Gemini 3.7 Flash pseudo-label: predicted IPA transcription from the same audio, used for teacher-label training experiments.
Gemini Pro	string	Gemini 3.1 Pro normalized stress-marked IPA pseudo-label from the same audio; null means the label did not pass validation.
name	string	Gold name text = what was actually spoken, with reference spellings where a spoken token matches the reference.
phonetic	string	Gold phonetic transcription of the reference name in a stress-marked notation (e.g. sen-TA-nee, HAR-nish).

Gemini IPA

Gemini Pro

Gemini Pro is a normalized, stress-marked IPA teacher label generated with Gemini 3.1 Pro. Null entries are rows whose output failed the validation gates.

Gemini IPA is a teacher pseudo-label generated from the clip audio with Gemini 3.7 Flash. It is provided for model-training experiments and is not an independent human-verified gold transcription.

How name is built

name is reconstructed from the spoken audio, using the directory reference only as a spelling oracle and safety net:

• A spoken token that fuzzy-matches a reference token (sequence-similarity ≥ 0.6) is corrected to the reference spelling — so a misrecognised Harnish → Harnisch, Fazio → D'Orazio, Santani → Centanni.
• A spoken token that matches no reference token is kept verbatim, so real names the (often unreliable) register got wrong are preserved: Jake Furst (register said Jacob), Maddy Eberhard (register said Madison), an inserted middle name like the Eileen in "Heather Eileen Harnish".
• Reference tokens the speaker never produced are dropped — an unpronounced middle name (Raymond) or initial (A.) does not appear.
• The surname (final register token) is kept as a backbone if the ASR merely failed to recognise it, and any trailing ASR hallucination (e.g. Aunt Emily) is discarded in that case.
• Apostrophes inside names (D'Orazio, O'Brien) are preserved; a D' prefix is never treated as an independent, droppable initial.
• A small corrections table overrides a handful of known-bad references (e.g. a stored initial I. that the full spoken name is Idan).
• Short filler / non-name words (articles, hesitations, digits, fragments) are discarded via a stop-list.
• If the clip has no usable ASR output, name falls back to the reference name (see the empty split).

Examples:

reference (name_hedb)	ASR raw	name	phonetic
Heather Harnisch	Heather Eileen Harnish.	Heather Eileen Harnisch	HAR-nish
Joseph Raymond Centanni	Joseph Centanni.	Joseph Centanni	sen-TA-nee
Nicholas A. Pelosi	Nicholas Pelosi.	Nicholas Pelosi	pe-LO-see
Dominic D'Orazio	Dominic Fazio.	Dominic D'Orazio	der-A-zee-oh
Jacob Furst	Jake Furst.	Jake Furst	FURST
Mary Golly Easterly	Mary Francis Easterly.	Mary Francis Easterly	EES-ter-lee
Angela Michelle Antonelli	Angela, Michelle, Aunt Emily.	Angela Michelle Antonelli	an-ti-NE-lee

Loading

from datasets import load_dataset

ds = load_dataset("Reza2kn/netjets-name-asr")
print(ds["train"][0])
# {'audio': {'array': ..., 'path': ..., 'sampling_rate': 16000},
#  'name': 'Heather Eileen Harnisch',
#  'phonetic': 'HAR-nish',
#  'ipa': '...',
#  'Gemini IPA': '...'}

Notes & limitations

• phonetic uses a custom stress-marked notation, not standard ARPABET — map it before using with ARPABET-based G2P tooling. It corresponds to the reference name, which may include syllables not present in a partial name (e.g. an unspoken middle name). Align at the token level if you need per-name phonetics.
• Audio is short, conversational, and often recorded over the phone in noisy conditions: expect accent, disfluency, and variable pronunciation.
• Speaker identities are not provided; this dataset must not be used for speaker attribution.

License

Provided for research / non-commercial use. Confirm licensing of the underlying recordings before any commercial application.