Audio Datasets
Collection
My training datasets for ASR fine tuning • 7 items • Updated
Datasets:
audio audio | id string | title string | recorded_at timestamp[s] | source_filename string | language string | speakers int64 | recording_context string | purposes list | purpose_explanation string | notes list | tags list | duration_s int64 | word_count int64 | transcript_raw string | asr_engine string | asr_settings dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
avn-0001 | Apartment remaining-fixes walkthrough | 2026-10-17T12:35:12 | 2026_10_17_15_35_12.mp3 | en | 1 | Walking room by room through the apartment with a phone recorder, listing the remaining fixes, purchases and projects. Unscripted, a single take. A toddler can be heard in the background at points. | [
"task_extraction",
"disfluency",
"long_form_task_recording"
] | A training and evaluation example for classifiers that pull tasks out of loosely structured free speech. The tasks are almost never stated as imperatives. They come through as observations ('the slats keep falling down'), plans ('we need to do one trip out there this week'), wishes ('I'd love to put up a couple of piec... | [
"Structure is spatial, not by priority: living room, kitchen, bathroom, patio (as an afterthought), bedroom, ensuite, home office, patio storage, then the bedroom again near the end. A task extractor has to cope with topics that come back later.",
"Tasks sit in mixed tiers that the speaker labels out loud: 'top p... | [
"home",
"to-do",
"walkthrough",
"shopping",
"smart-home",
"networking",
"long-form"
] | 1,339 | 3,537 | Okay, today's October 17th. List, uh, of fixes. M. What's left to do for the apartment. I'm gonna just be walking around room by room and making notes for what we need and what's. What's uh, delaying stuff. Little Ezra here might be making some contributions of his own. Um, Ezra, what do we need to get done? Okay, in t... | assemblyai/best | {
"disfluencies": true,
"language_code": "en_us"
} |
Annotated Voice Notes
A small, hand-picked set of real voice notes. It is kept small on purpose. Each note has its audio, a raw AssemblyAI transcript, and a written annotation that explains why it was included and what it is good for.
This is not a bulk corpus. Every recording was chosen because it shows something specific, such as tasks buried in rambling speech, heavy disfluency, or a long take. The annotation says what that is.
Fields
| Field | Description |
|---|---|
audio |
The recording, in its original format (MP3) |
id |
Stable sample id (avn-NNNN) |
title |
Short description |
recorded_at |
ISO 8601 with offset |
recording_context |
How and where it was recorded |
purposes |
Machine-readable list of the uses it was included for |
purpose_explanation |
Why it is in the dataset |
notes |
Observations on structure, disfluencies and likely ASR errors |
tags |
Topic tags |
duration_s, word_count |
From the ASR output |
transcript_raw |
AssemblyAI output, uncorrected, with disfluencies kept |
asr_engine, asr_settings |
Model and settings that produced the transcript |
Transcripts are raw. They are not corrected against the audio, so known and
suspected ASR errors are listed in notes and left in the text.
Samples
| id | Title | Length | Purposes |
|---|---|---|---|
| avn-0001 | Apartment remaining-fixes walkthrough | 22:19 | task extraction, disfluency, long-form |
Purposes vocabulary
task_extraction: tasks expressed loosely, not as a clean listdisfluency: fillers, false starts, self-corrections, restartslong_form_task_recording: several minutes of continuous task-bearing speech
- Downloads last month
- 48