Dataset Viewer
Auto-converted to Parquet Duplicate
audio
audioduration (s)
2.57
9.67
id
stringclasses
6 values
text
stringclasses
6 values
duration_s
float64
2.57
9.67
categories
listlengths
1
1
target_terms
listlengths
0
0
prompt_id
stringclasses
6 values
prompt_source
stringclasses
1 value
mode
stringclasses
1 value
edited
bool
1 class
device
stringclasses
1 value
environment
stringclasses
1 value
sample_rate
int64
16k
48k
recorded_at
stringclasses
6 values
microphone
stringclasses
1 value
prompt_origin
stringclasses
1 value
hebrew_pronunciation
stringclasses
0 values
variant_groups
listlengths
0
0
variant_spelling
stringclasses
0 values
speaker_id
stringdate
8992-01-01 00:00:00
8992-01-01 00:00:00
conditions
listlengths
0
0
imported_from
stringclasses
0 values
original_text
stringclasses
0 values
text_before_normalization
stringclasses
0 values
text_he
stringclasses
6 values
noise_level
stringclasses
0 values
background_speech_languages
listlengths
0
0
notes
stringclasses
0 values
model_hypothesis
stringclasses
0 values
asr_model
stringclasses
0 values
speaker
stringclasses
1 value
speaker_gender
stringclasses
1 value
noise_flags
listlengths
0
0
clipped
bool
1 class
task
stringclasses
1 value
spoken_text
stringclasses
6 values
aside_text
stringclasses
0 values
aside_position
stringclasses
0 values
verbatim_text
stringclasses
0 values
prompt_categories
listlengths
source_prompt_id
stringclasses
0 values
prompt_kind
stringclasses
2 values
prompt_note
stringclasses
0 values
term_language
stringclasses
0 values
paragraph_text
stringclasses
0 values
contributor
stringclasses
1 value
bit_depth
int64
16
24
audio_format
stringclasses
2 values
native_rate
int64
48k
48k
⌀
capture_mode
stringclasses
1 value
tts_eligible
bool
2 classes
tts_split
stringclasses
0 values
transcript_status
stringclasses
1 value
prompt_set
stringclasses
0 values
topic
stringclasses
0 values
style
stringclasses
0 values
style_instruction
stringclasses
0 values
written_form_uncertain
bool
0 classes
session_id
stringclasses
2 values
room_tone_id
stringclasses
0 values
lead_trim_s
float64
0
0
⌀
peak_dbfs
float64
-6.9
-2.8
⌀
floor_dbfs
float64
-52
-33.5
⌀
speech_dbfs
float64
-23.4
-14.3
⌀
snr_db
float64
19.2
28.6
⌀
clipped_samples
int64
0
0
⌀
edge_head_dbfs
float64
-41.9
-22.4
⌀
edge_tail_dbfs
float64
-39.3
-22.2
⌀
browser_processing
bool
1 class
room
stringclasses
0 values
priority_word
stringclasses
0 values
location
stringclasses
0 values
location_name
stringclasses
0 values
location_note
stringclasses
0 values
latitude
float64
longitude
float64
location_accuracy_m
float64
text_full
stringclasses
0 values
fallback_ok
bool
0 classes
sentence_id
stringclasses
6 values
pronunciation
stringclasses
1 value
extra
stringclasses
1 value
20261021T115048Z_3rp02o
I think the report is late.
2.574
[ "false-starts" ]
[]
false-starts_b30a422770
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
16,000
2026-10-21T11:50:48.473Z
internal
generated
null
[]
null
8992
[]
null
null
null
I think the report is late.
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
I think the the report is late.
null
null
null
null
null
repeat-word
null
null
null
Daniel Rosehill
16
wav
null
null
null
null
script
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
p_ca485a5031
[]
null
20261021T115148Z_jje4tg
Can you send me the latest version of the contract?
6.45
[ "false-starts" ]
[]
false-starts_d1e7b791ac
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
16,000
2026-10-21T11:51:48.538Z
internal
generated
null
[]
null
8992
[]
null
null
null
Can you send me the latest version of the contract?
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
Can you send me the the latest version of the contract?
null
null
null
null
null
repeat-word
null
null
null
Daniel Rosehill
16
wav
null
null
null
null
script
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
p_ba5f86e8de
[]
null
20261021T120237Z_1am0ty
I wanted to check whether you got my email from yesterday.
9.67
[ "false-starts" ]
[]
false-starts_81a345ef06
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
16,000
2026-10-21T12:02:37.820Z
internal
generated
null
[]
null
8992
[]
null
null
null
I wanted to check whether you got my email from yesterday.
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
I I wanted to check whether you got my email from yesterday.
null
null
null
null
null
repeat-word
null
null
null
Daniel Rosehill
16
wav
null
null
null
null
script
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
null
p_c0f5c0743b
[]
null
20261030T112514Z_uhyzzn
The meeting is on Thursday at 3:30 in the small room.
5.45
[ "false-starts" ]
[]
false-starts_6a156cb54b
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
48,000
2026-10-30T11:25:14.681Z
internal
generated
null
[]
null
8992
[]
null
null
null
The meeting is on Thursday at 3:30 in the small room.
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
The meeting is on Thursday at at 3:30 in the small room.
null
null
null
null
null
repeat-word
null
null
null
Daniel Rosehill
24
flac
48,000
stt
false
null
script
null
null
null
null
null
20261030T105428Z_nk6jic
null
0
-3
-33.5
-14.3
19.2
0
-22.4
-22.7
false
null
null
null
null
null
null
null
null
null
null
p_1f94dfdfcd
[]
null
20261030T112519Z_eczsq2
We need to order more printer paper before Monday.
3.834
[ "false-starts" ]
[]
false-starts_e8f472e211
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
48,000
2026-10-30T11:25:19.476Z
internal
generated
null
[]
null
8992
[]
null
null
null
We need to order more printer paper before Monday.
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
We need to to order more printer paper before Monday.
null
null
null
null
null
repeat-word
null
null
null
Daniel Rosehill
24
flac
48,000
stt
false
null
script
null
null
null
null
null
20261030T105428Z_nk6jic
null
0
-2.8
-34.8
-15
19.8
0
-23.8
-22.2
false
null
null
null
null
null
null
null
null
null
null
p_b2b03bf295
[]
null
20261119T205119Z_a9vb0d
Remind me to water the plants tonight.
4.637
[ "false-starts" ]
[]
false-starts_b6ed17eec2
generated:claude-self-written
bank
false
OnePlus Nord 3 5G
quiet-indoor
48,000
2026-11-19T20:51:19.873Z
internal
generated
null
[]
null
8992
[]
null
null
null
Remind me to water the plants tonight.
null
[]
null
null
null
Daniel Rosehill
male
[]
false
false-starts
I was going to, Remind me to water the plants tonight.
null
null
null
null
null
abandoned-start
null
null
null
Daniel Rosehill
16
flac
48,000
stt
true
null
script
null
null
null
null
null
20261119T205054Z_jo79vc
null
0
-6.9
-52
-23.4
28.6
0
-41.9
-39.3
false
null
null
null
null
null
null
null
null
null
null
p_016a573f09
[]
{"mic_link":null,"bt_codec":null,"mic_name":"CPH2493","mic_usb_vendor_id":null,"mic_usb_product_id":null,"mic_type":"builtin_mic","mic_chain":null,"activity":"seated","background_speech_kinds":[],"pronunciation_variants":[],"ui_note":null,"source_url":null,"check_verdict":null,"check_errors":null,"check_error_types":nu...

False-Start Sentences

Type: Audio Role: Split-out Subset of: Read-Aloud-Voice-Sentences

Audio dataset, split-out. Rows copied from the master, danielrosehill/Read-Aloud-Voice-Sentences, so one task can be browsed alone. Training reads the master, not this copy. Every audio dataset as one tree, with hours: the audio index. How every dataset fits together: the dataset index.

Experimental. One speaker reading scripted false starts: a word or phrase said twice ("the the report", "can you, can you"), a mid-sentence restart ("We should — we need to call him") and an abandoned start followed by a fresh sentence. spoken_text is the script as read; text is the transcript with the abandoned words dropped. prompt_kind is repeat-word, repeat-phrase, restart, abandoned-start or control; controls are real doubled words ("had had", "very, very") that the transcript keeps. Content-changing corrections are in Self-Correction-Sentences, fillers in Disfluent-Read-Sentences.

Website: hebrish.app: get the Hebrish keyboard for Android and the desktop apps, see the results, and find every model, dataset and demo in one place.

Two tabs. sentences lists every ground-truth sentence with a takes count, so takes = 0 shows what has not been recorded. takes has one row per recording with its audio, joined on sentence_id; a sentence can have several takes. load_dataset("Hebrish/False-Start-Sentences", "sentences") reads the text; the default config reads the takes.

Takes come from danielrosehill/Read-Aloud-Voice-Sentences, the full dataset, with the same ids and train/test split, by scripts/sync_subsets.py in the project repository. Do not edit this dataset directly. New takes append and only the sentence counts are rewritten.

Speaker: Daniel Rosehill, the dataset's author, who was born in Ireland and speaks English with an Irish accent, and who lives in Jerusalem. Every recording is his voice. Deliberate variation: he sometimes changes his speaking speed on purpose (fast and rushed, or slow) and sometimes says a word the way people commonly mispronounce it, to make the models robust to real speech. The transcript is always the sentence as written, so a take whose speed or pronunciation sounds off against its card is usually intentional, not a recording error.

All audio was recorded on a OnePlus Nord 3 5G phone using its internal microphone, with browser echo cancellation, noise suppression and auto gain turned off.

Layout

sentences/train.parquet  one row per sentence: `id`, `text`, `takes`, `prompt_ids`, `served` and the sentence columns
data/train-*.parquet   one row per take: `audio` (the FLAC or WAV file, embedded) and every column
meta/train.jsonl       index: the same rows without audio, plus `shard`, `audio_path` and the audio's byte range
meta/schema.json       column types
data/test-*.parquet, meta/test.jsonl   held-out set, never trained on

Audio is 48 kHz mono 24-bit FLAC for rows from 2026-10-22 and 16 kHz mono 16-bit WAV before.

The test split is term-disjoint: a held-out term never appears in train. load_dataset("Hebrish/False-Start-Sentences") works directly.

Columns

A take's sentence_id is its row's id in sentences. A sentence with served: false is no longer offered by the app; it is kept because a take used it. The columns are the same as in the main dataset: text (ground truth, punctuated and cased like Parakeet v2 output), text_he (the same transcript with its Hebrew terms in Hebrew script, for Hebrew-capable models), categories, target_terms, prompt_source, prompt_origin (manual / dataset / generated), edited, duration_s, environment, conditions, device, speaker, speaker_gender, contributor, hebrew_pronunciation, variant_groups, variant_spelling and recorded_at. The main dataset's card describes them.

Licence

CC-BY-4.0, with sentence-text attribution per prompt_source as described in the main dataset's licence section.

Downloads last month
129