Dataset Viewer
Auto-converted to Parquet Duplicate
audio
audio
id
string
room
string
conditions
list
notes
string
source
string
duration_s
float64
sample_rate
int64
bit_depth
int64
audio_format
string
native_rate
int64
peak_dbfs
float64
floor_dbfs
float64
clipped_samples
int64
browser_processing
bool
device
string
microphone
string
input_label
string
session_id
string
contributor
null
recorded_at
string
mic_name
string
mic_usb_vendor_id
string
mic_usb_product_id
string
mic_type
string
mic_chain
null
20261022T141053Z_bfbk3r
living-room
[ "ac-on" ]
null
room-tone-screen
29.498
48,000
24
flac
48,000
-37.6
-55.1
0
false
OnePlus Nord 3 5G
internal
null
20261022T132308Z_dllnyx
null
2026-10-22T14:10:53.171Z
null
null
null
null
null
20261022T141140Z_oq4yfk
home-office
[ "ac-on" ]
null
room-tone-screen
29.295
48,000
24
flac
48,000
-45.2
-62
0
false
OnePlus Nord 3 5G
internal
null
20261022T132308Z_dllnyx
null
2026-10-22T14:11:40.691Z
null
null
null
null
null
20261112T200953Z_vnayc9
home-office
[ "ac-on", "appliance-hum" ]
AC on and a white noise machine running. Recorded from the raw Boya BY-CM1 (not the ASR Mic chain) with pw-record on the desktop, by Claude at Daniel's request; 30 s recorded, trimmed to the clean first 13.7 s because speech started at 16 s.
desktop-pw-record
13.7
48,000
24
flac
48,000
-30.9
-61
0
false
Desktop workstation
BY-CM1
BY-CM1 Analog Stereo
null
null
2026-11-12T20:09:53Z
BY-CM1
ff01
000b
usb
null

Room tone annotations for ASR correction

Room tone: recordings of silence in a labelled room, made with the same phone, microphone and settings as the Hebrish read-speech recordings. Each recording is tagged with the room, the conditions at the time (window open, AC, appliance hum, traffic …) and an optional note.

Uses: noise profiles for denoising, and matching speech takes to the room they were recorded in. The speech corpora carry the same room value as this dataset, and a room_tone_id when a room tone was recorded in that room in the same session.

Format

data/train/audio/<id>.flac: 48 kHz, 24-bit, mono FLAC, 10 to 60 seconds, untrimmed. Browser echo cancellation, noise suppression and auto gain are requested off (browser_processing records whether the device honoured that). data/train/metadata.jsonl has one row per recording.

Column Meaning
room home-office, living-room, kitchen, bedroom (null on early session recordings)
conditions window-open, door-open, ac-on, fan-on, heating-on, appliance-hum, computer-fan, traffic-outside, people-elsewhere, rain, night
notes Free text
source room-tone-screen, or session for room tone captured at the start of a voice-clone session
duration_s, sample_rate, bit_depth, native_rate Audio facts
peak_dbfs, floor_dbfs, clipped_samples Level figures
device, microphone, input_label, session_id, recorded_at Capture context

Rows are appended by the recording app and never rewritten.

from datasets import load_dataset
ds = load_dataset("Hebrish/Room-Tone-Annotations", split="train")
Downloads last month
150

Space using Hebrish/Room-Tone-Annotations 1