---
title: "How to test unsupervised diarization vs YouTube ASR on hula-hoop clips"
site: "transcribeall.io"
kind: "article"
author: "Piper Bowen"
date: "November 12, 2026"
canonical: "https://transcribeall.io/blog/how-to-test-unsupervised-diarization-vs-youtube-asr-on-hula-hoop-clips.php"
markdown: "https://transcribeall.io/blog/how-to-test-unsupervised-diarization-vs-youtube-asr-on-hula-hoop-clips.php/index.md"
description: "How unsupervised diarization clusters 20 hula-hoop clips against YouTube ASR word timing, accepting labels only when DER passes the 20% threshold gate."
---

# How to test unsupervised diarization vs YouTube ASR on hula-hoop clips

Piper Bowen · November 12, 2026

> How unsupervised diarization clusters 20 hula-hoop clips against YouTube ASR word timing, accepting labels only when DER passes the 20% threshold gate.

This guide delivers a matched-20-clip comparison of unsupervised diarization clustering versus YouTube ASR word timing for hula-hoop videos.

The clustering result is accepted for speaker labels only when DER passes the pre-registered 20% threshold gate.

![This guide delivers a matched-20-clip comparison of unsupervised — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-ai/how-to-test-unsupervised-diarization-vs-ai-b171d54e.jpg)

## How hula-hoop audio breaks diarization

Running VAD first on every hula-hoop clip is non-negotiable, because hoop impacts and background music are acoustically non-speech yet spectrally broadband. If VAD misclassifies these transients as voice, the downstream clustering step will fabricate speaker turns before any ASR comparison begins, corrupting the entire diarization chain. A conservative energy-threshold VAD that suppresses short-duration transients below a rolling median is therefore the first gate; only segments passing this filter are forwarded to embedding extraction.

Speaker embeddings must be extracted exclusively from the VAD-filtered voice-only segments, then clustered with spectral clustering. The fixed speaker count is recorded at this stage, because the grounded telephone benchmark assumes a known two-speaker setting, whereas a music-heavy hula-hoop clip may contain zero, one, or multiple voiceover artists. Clustering is performed on the embedding matrix without any pre-trained speaker model, preserving the unsupervised requirement and avoiding leakage from external datasets.

After clustering, YouTube ASR word timestamps are aligned to the diarization turns. Alignment is done by mapping each ASR token to the nearest cluster centroid in time, then merging contiguous tokens that share the same cluster label. This step converts raw word-level timestamps into speaker-attributed turns, enabling direct computation of WER for ASR and DER for clustering on the same 20-clip corpus.

Pre-register the hula-hoop DER gate before any run; the 0.9% telephone-conversation result is not a hula-hoop gate, because no grounded hula-hoop DER exists. Use clustering for speaker labels only if the matched 20-clip DER passes that pre-registered gate. If it fails, keep YouTube ASR for word timing only and flag who-spoke-when for manual review rather than crediting clustering with unshown accuracy.

No entry in the table supplies a hula-hoop DER or a YouTube-ASR diarization accuracy. Any claim that these figures transfer directly to hula-hoop clips would exceed the available sources. The only defensible path is the pre-registered 20-clip test where clustering must pass its DER gate before speaker labels are trusted.

![How hula-hoop audio breaks diarization — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-ai/how-to-test-unsupervised-diarization-vs-ai-15181001.jpg)

## Unsupervised clustering vs YouTube ASR

No winner is declared before the matched 20-clip test runs. Unsupervised clustering is the candidate for who-spoke-when only if its DER passes the pre-registered gate; YouTube ASR is used only as a word-timing baseline. The table below summarizes the functional roles and scoring criteria on the same hand-corrected hula-hoop excerpt.

| Method | Produces | Scoring Metric | Diarization Winner? |
| --- | --- | --- | --- |
| Unsupervised Clustering | Speaker turns | DER (Diarization Error Rate) | Yes, if DER passes pre-registered gate |
| YouTube ASR | Word timings | WER (Word Error Rate) | No, lacks speaker labels |

Score both on the same 3-minute hand-corrected hula-hoop excerpt: compute DER for clustering and WER for YouTube ASR, and never declare a diarization winner from a WER number. A DER value must be calculated against ground-truth speaker labels, while WER reflects transcription accuracy only. If the deliverable is searchable text without who-said-what, choose YouTube ASR, as it provides usable word sequences even without speaker attribution.

Unsupervised clustering is the winner for diarization if it passes your hula-hoop reference score. YouTube ASR is the loser for diarization when it has no speaker labels. The clustering method must be evaluated on its ability to assign consistent speaker identities across turns, not on transcription fidelity. A pre-registered DER threshold—derived from your hula-hoop test set—must be met before clustering is permitted to supply speaker labels.

Use clustering for speaker labels only if DER passes your pre-registered gate. If it fails, fall back to YouTube ASR for word timing and accept the loss of speaker attribution. This rule ensures that no method is credited with diarization capability unless it explicitly demonstrates speaker discrimination on the hula-hoop test set.

![Unsupervised clustering vs YouTube ASR — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-pixabay/how-to-test-unsupervised-diarization-vs-71166804.jpg)

## Hula-hoop run costs and DER units

LecSync lists diarization inside its base rate of $0.006 per audio minute, which converts to $0.36 per audio hour. The source note confirms that diarization is included in this figure; no additional streaming fee applies. AssemblyAI charges an additional $0.12 per hour for streaming diarization as of October 6, 2026, on top of the base ASR mode. This per-hour surcharge must never be added to the LecSync per-minute figure without first converting LecSync to a per-hour baseline, ensuring the two cost structures are compared on the same audio-hour denominator.

DER and WER cannot be compared directly because they measure different error types. Diarization Error Rate (DER) counts speaker-attribution errors and missed-speech segments, while Word Error Rate (WER) counts word-level insertions, deletions, and substitutions. A hula-hoop clip that contains background music may show an acceptable WER because the ASR output captures the beat or melody as text, but the same clip could produce a high DER if the system fails to segment the voiceover from the music. This denominator mismatch means that a low WER does not guarantee accurate speaker labels, and vice versa.

![Hula-hoop run costs and DER units — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-pixabay/how-to-test-unsupervised-diarization-vs-6ac3eb68.jpg)

## Where the telephone benchmark stops

Where the telephone benchmark stops

This section alone lists hula-hoop edge cases where telephone-derived DER stops transferring: music-only, single voiceover, overlapping impacts, and more than two speakers. The telephone-derived 0.9% DER figure from the unsupervised diarization paper (people.csail.mit.edu) is a two-speaker, clean-channel result; it does not survive the acoustic conditions found in hula-hoop videos.

A music-only intro contains no speaker to cluster, so the unsupervised clustering rule breaks by definition. If the clip is entirely instrumental, VAD returns no voice segments, and the DER gate cannot be evaluated. A single voiceover over music, by contrast, can still pass if VAD cleanly separates the voice from the background track; the clustering step then sees one speaker and trivially meets the DER threshold.

Overlapping hoop impacts and speech break the rule unless source separation precedes clustering. Hoop transients are broadband and non-speech, yet they can be misclassified as voice by VAD, injecting false speaker segments that inflate DER. Frontiers reports that unsupervised methods yield higher error than supervised ones (frontiersin.org); this finding is used here to require a hand-checked hula-hoop subset before trusting any unsupervised pass, not as a substitute for it.

More than two speakers breaks the rule until retested. The 0.9% DER figure assumes exactly two speakers; adding a third or fourth voice introduces permutation errors that unsupervised clustering cannot resolve without additional constraints. Each additional speaker should be treated as a new experiment, not an extrapolation.

| Edge case | Rule status | Required precondition |
| --- | --- | --- |
| Music-only intro | Breaks | None; no speaker exists |
| Single voiceover over music | Wins if VAD clean | VAD must separate voice from music |
| Overlapping hoop impacts and speech | Breaks | Source separation before clustering |
| More than two speakers | Breaks | Retest as new experiment |

![Where the telephone benchmark stops — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-pixabay/how-to-test-unsupervised-diarization-vs-075e8cbc.jpg)

## Worked check: 20 hula-hoop clips

For the 2026 hula-hoop video test, the only defensible comparison is a matched 20-clip worksheet because no grounded source supplies a hula-hoop or YouTube-ASR diarization accuracy number. This section alone applies a 20-clip by five-minute worksheet to compute one hundred minutes, one point six six six seven audio hours, zero point six zero dollars LecSync, zero point twenty dollars AssemblyAI add-on, and the rework-cost comparison. The inputs are twenty hula-hoop clips multiplied by five minutes each, yielding one hundred minutes or one point six six six seven audio hours. LecSync runs at zero point zero zero six dollars per audio minute, so the total is zero point sixty dollars. AssemblyAI charges a diarization add-on of zero point twelve dollars per hour, which adds zero point twenty dollars for the one point six six six seven hours.

The worksheet records clip ID, duration, VAD speech minutes, reference speaker count, DER numerator and denominator, WER after stripping speaker labels, and chosen pipeline. Keep any telephone-derived DER in a separate not-hula-hoop column. Pre-register the hula-hoop DER gate before any run; the 0.9% telephone baseline is not that gate. If clustering fails the gate, keep YouTube ASR for word timing only and flag speaker attribution for manual review; do not treat YouTube ASR as supplying speaker labels.

The available sources confirms that the zero point nine percent DER figure applies to two-speaker telephone conversations and is not validated for hula-hoop videos. The LecSync pricing of zero point zero zero six dollars per audio minute and the AssemblyAI add-on of zero point twelve dollars per hour are as of October 6, 2026, per lecsync.com. The dcaulley GitHub recipe for compiling YouTube channel URLs is noted, but it is not a hula-hoop or YouTube-ASR diarization accuracy source. The worksheet must be run with VAD first, because hoop impacts and background music are acoustically non-speech yet spectrally broadband, and misclassifying these transients as voice will fabricate speaker embeddings.

The unsupervised clustering pipeline uses VAD, speaker embeddings, spectral clustering, and ASR alignment, as defined by this section alone. The winner is unsupervised clustering for who-spoke-when in hula-hoop videos when the DER passes the pre-registered gate. The rework-cost comparison is computed by multiplying the total audio hours by the respective rates and adding any add-ons. The worksheet is designed to be copy-usable, with columns that allow direct comparison between unsupervised clustering and YouTube ASR speaker labels. No other section may restate or re-explain these claims.

## Four if/then rules for hula-hoop clips

If a hula-hoop video requires who-spoke-when labels and contains music or hoop-impact transients, route it to the VAD-plus-spectral-clustering path. If the clip only needs searchable words and no speaker attribution, route it to YouTube ASR. This first decision prevents downstream clustering from being polluted by non-speech broadband noise that mimics voice.

When the clip has more than two speakers or overlapping voiceover, do not accept a telephone-condition DER as a passing score. Run the 20-clip worksheet or send the file to manual review. If YouTube ASR output contains no speaker labels, never treat it as diarization; use it only as word-timing input to the clustering path. This rule keeps the two technologies from being conflated in the final report.

If the 20-clip DER miss rate exceeds the pre-registered gate, the unsupervised clustering result is rejected for speaker-labeling purposes. In that case, fall back to YouTube ASR for word-timing only and mark the clip as “manual review required.” The gate is set before any audio is processed, so no post-hoc adjustment is allowed.

For every hula-hoop clip, record which of the four branches was taken and the raw DER or WER value. This creates an audit trail that can be checked by any third party without re-running the models. The worksheet itself is the only artifact that ties the clip to the chosen path, so keep it alongside the final output.

For each clip, record which branch was taken and the raw DER or WER value so a third party can audit the choice without rerunning the models. Keep the worksheet with the final output as the artifact tying each clip to its chosen path.

## Frequently Asked Questions

**What is the maximum DER allowed for the clustering result to be accepted for speaker labels?**

The clustering result is accepted for speaker labels only when DER passes the pre-registered 20% threshold gate.

**Why must VAD be run before diarization on hula-hoop clips?**

Hoop impacts and background music are acoustically non-speech yet spectrally broadband, and if VAD misclassifies these transients as voice, the downstream clustering step will fabricate speaker turns before any ASR comparison begins.

**What type of VAD is recommended for filtering hula-hoop audio?**

A conservative energy-threshold VAD that suppresses short-duration transients below a rolling median is the first gate.

**From which segments are speaker embeddings extracted?**

Speaker embeddings must be extracted exclusively from the VAD-filtered voice-only segments.

**What clustering method is used on the embedding matrix?**

Clustering is performed on the embedding matrix with spectral clustering without any pre-trained speaker model.

**How is the speaker count determined for the hula-hoop clips?**

The fixed speaker count is recorded at the clustering stage because a music-heavy hula-hoop clip may contain zero, one, or multiple voiceover artists.

## Quick answers

| What primary comparison does the guide deliver for hula-hoop videos? | The guide delivers a matched-20-clip comparison of unsupervised diarization clustering versus YouTube ASR word timing for hula-hoop videos. |
| --- | --- |
| At what DER threshold is the clustering result accepted for speaker labels? | The clustering result is accepted for speaker labels only when DER passes the pre-registered 20% threshold gate. |
| Why is VAD run first on every hula-hoop clip? | Running VAD first on every hula-hoop clip is non-negotiable, because hoop impacts and background music are acoustically non-speech yet spectrally broadband. |
| What type of VAD is recommended to avoid misclassifying transients as voice? | A conservative energy-threshold VAD that suppresses short-duration transients below a rolling median is therefore the first gate. |
| How is clustering performed to maintain the unsupervised requirement? | Clustering is performed on the embedding matrix without any pre-trained speaker model, preserving the unsupervised requirement and avoiding leakage from external datasets. |

Also worth reading: **Transform your audio and video files into accurate text transcripts with AI**: [Transform your audio and video](https://transcribeall.io/blog/transform-your-audio-and-video-files-into-accurate-text-transcripts-with-ai.php) · **pyannote.audio 3.1: verify a 0.7→0.6 clustering threshold swap on CHiME-6**: [pyannote.audio 3.1: verify a 0.7→0.6](https://transcribeall.io/blog/pyannoteaudio-31-verify-a-0706-clustering-threshold-swap-on-chime-6.php) · **Diarization Cuts Podcast WER by 18.4%: Pre vs Post ASR**: [Diarization Cuts Podcast WER by](https://transcribeall.io/blog/diarization-cuts-podcast-wer-by-184-pre-vs-post-asr.php)

### Related reading

- [Podcast transcript editing time: 15% Diarization Error Rate (DER) vs VAD tuning](https://transcribeall.io/blog/podcast-transcript-editing-time-15-diarization-error-rate-der-vs-vad-tuning.php)
- [Launching a Podcast in 2026: Deploy Diarization at 0.8 to Separate Hosts](https://transcribeall.io/blog/launching-a-podcast-in-2026-deploy-diarization-at-08-to-separate-hosts.php)
- [Podcast Ad Transcription WER: Diarization When Overlap Hits 25%](https://transcribeall.io/blog/podcast-ad-transcription-wer-diarization-when-overlap-hits-25.php)
- [Free Audio to Text Converter: 30-Minute Upload Limit vs Speaker Diarization](https://transcribeall.io/blog/free-audio-to-text-converter-30-minute-upload-limit-vs-speaker-diarization.php)
- [Transcribing Lateral Podcast Episode 68: Diarization vs Manual Labels for 3+ Voices](https://transcribeall.io/blog/transcribing-lateral-podcast-episode-68-diarization-vs-manual-labels-for-3-voices.php)
- [Cisco Support Services 2026: 10% WER Threshold vs Speaker Diarization](https://transcribeall.io/blog/cisco-support-services-2026-10-wer-threshold-vs-speaker-diarization.php)

### Latest

- [Apple One after the 2026 price hike: run this break-even test before canceling](https://transcribeall.io/blog/apple-one-after-the-2026-price-hike-run-this-break-even-test-before-canceling.php)
- [Hark Pro computer use: Handoff² 2x faster finish 82% of tasks](https://transcribeall.io/blog/hark-pro-computer-use-handoff-2x-faster-finish-82-of-tasks.php)
- [African speech recognition: FLEURS 102, Whisper large-v3 vs Meta MMS-1B, lower...](https://transcribeall.io/blog/african-speech-recognition-fleurs-102-whisper-large-v3-vs-meta-mms-1b-lower-wer.php)

Canonical: https://transcribeall.io/blog/how-to-test-unsupervised-diarization-vs-youtube-asr-on-hula-hoop-clips.php
Markdown: https://transcribeall.io/blog/how-to-test-unsupervised-diarization-vs-youtube-asr-on-hula-hoop-clips.php/index.md

```json
{"@context":"https://schema.org","@type":"BlogPosting","headline":"How to test unsupervised diarization vs YouTube ASR on hula-hoop clips","url":"https://transcribeall.io/blog/how-to-test-unsupervised-diarization-vs-youtube-asr-on-hula-hoop-clips.php","description":"How unsupervised diarization clusters 20 hula-hoop clips against YouTube ASR word timing, accepting labels only when DER passes the 20% threshold gate.","datePublished":"November 12, 2026","author":{"@type":"Person","name":"Piper Bowen"},"articleBody":"This guide delivers a matched-20-clip comparison of unsupervised diarization clustering versus YouTube ASR word timing for hula-hoop videos. The clustering result is accepted for speaker labels only when DER passes the pre-registered 20% threshold gate. ![This guide delivers a matched-20-clip comparison of unsupervised — How to test unsupervised diarization vs](https://static.mm-ais.com/article-images-ai/how-to-test-unsupervised-diarization-vs-ai-b171d54e.jpg) ## How hul..."}
```
