Audio level & timing
Signal-level changes in selected recording windows. Speech boundaries and anonymous voice groups estimated with Silero VAD and WeSpeaker embeddings.
We were as excited as you about Dessa. So we built a concept to explore how audio could be tested with QualityMax.
Listen to the published conversation, explore who speaks when, and inspect measured changes in the recording. Some questions still need a human ear.
A QualityMax research prototype · Recorded audio, analyzed offline
Quieter selected window
Louder selected window
The original audio from the linked X recording, at its original level. Playback follows the speaker timeline below.
Source: linked X recording / Block’s Dessa conversation. Anonymous speaker labels are model estimates. Source attribution ↗
Text from Block’s published conversation captions, aligned to this recording. Voice numbers follow the publisher’s labels, separately from our estimated speaker clusters. Captions may contain errors and omit non-speech sounds.
Voice 1: Hey, how you doing?
Voice 2: I'm doing well.
Voice 2: How are you doing with the winter
Voice 2: stuff?
Voice 1: Winter is coming for sure, but you
Voice 1: know
Voice 2: Is that kind of a sore spot
Voice 2: this year?
Voice 1: Yeah, it's gonna be cold that I'm
Voice 1: not a fan of the snow I
Voice 1: prefer to have sun on my skin.
Voice 1: You know what I'm saying?
Voice 2: Yeah, I can see that.
Voice 1: Well, anyway, I've had such an exciting
Voice 1: day at work today.
Voice 2: Oh, tell us.
Voice 1: Yeah, we the team
Voice 1: We've decided to release the speech-to
Voice 1: -speech AI model we've built for the
Voice 1: past year and
Voice 2: Whoa.
Voice 2: Whoa.
Voice 2: So I have been watching your journey
Voice 2: of creating this.
Voice 1: What did you think of it?
Voice 1: Are you excited about it as much
Voice 2: I was so blown away.
Voice 1: as I am?
Voice 2: I am so impressed, to be honest.
Voice 1: Yeah, it's so expressive, right?
Voice 2: It is so expressive, it, it's
Voice 2: conversational it's
Voice 1: It's it's it's very conversational.
Voice 1: It can laugh.
Voice 1: It can understand you and your emotions
Voice 1: much better than anything I've seen out
Voice 1: there.
Voice 2: I haven't really seen
Voice 2: anything like that either but it's it's
Voice 2: pretty much like human level.
Voice 1: Yeah, it's a bit scary, to be
Voice 1: honest,
Voice 1: but
Voice 1: but I'm so excited to see where
Voice 1: it would get us.
Voice 2: I'm
Voice 2: so
Voice 2: excited for it to be in the
Voice 2: world because it's just amazing we've had
Voice 1: I know.
Voice 2: so
Voice 1: So I totally agree.
Voice 2: many conversations about it.
Voice 1: What are you excited for?
Voice 1: What's your vision for where this AI
Voice 1: is going to be?
Voice 2: I, well, okay, so the one thing
Voice 2: that I've been kind of
Voice 2: about is advertising companies creating
Voice 2: their own AI bot.
Voice 1: No, fuck advertising.
Voice 1: If you were able to talk to
Voice 1: this AI, what will you tell it?
Voice 2: Oh, I would ask it so
Voice 1: What would you ask it?
Voice 2: many questions so if it could be
Voice 1: Like what?
Voice 2: talking it I don't know off the
Voice 2: bat I would ask
Voice 2: it for all the secrets
Voice 2: like
Voice 1: Like what secrets?
Voice 2: I would honestly ask it like what
Voice 1: Sorry, you're a bit far from the
Voice 2: is
Voice 2: consciousness I
Voice 1: mic.
Voice 1: Can you come closer so I can
Voice 1: hear you?
Voice 2: would be like what is consciousness is
Voice 2: it something that's
Voice 2: programmed into us is it something that
Voice 2: can be expanded on is it something
Voice 2: that's interdimensional
Voice 1: Those are all good questions.
Voice 1: But I can actually, I can't really
Voice 1: hear you.
Voice 1: Maybe you should cuff up, cough a
Voice 1: little bit or something so I can
Voice 1: hear your voice better.
Voice 2: Sorry.
Voice 1: Yeah, I think that's better.
Voice 1: Can you talk again?
Voice 2: Yeah, sure,
Voice 1: Yeah, yeah, that's much clearer.
Voice 1: Okay, so what were you saying?
Voice 2: I would ask I would like start off
Voice 2: with
Voice 2: two plus two questions and then like
Voice 1: Oh
Voice 1: Like whether it can emergently through
Voice 2: yeah
Voice 1: learning all this audio be able to
Voice 1: count
Voice 1: Do you think it'd be able to
Voice 1: do that?
Voice 2: Oh yeah, for sure.
A request to improve audibility creates a useful question: does the recording get louder afterward? Start with the sound, then inspect the selected windows.
These windows come from a mixed recording. Gain, speech content, and speaker attribution affect the comparison. This is not a pass/fail verdict on Dessa.
A short speech window before the request. This anchors the first comparison.
Open “Talking about the release” and seek to about 2:02. Its timeline starts approximately 0.55 s earlier than the X recording analyzed here.
The analyzed recording was shared on X. Block’s release page publishes the corresponding conversation and synchronized speaker tracks.
X media metadata credits @blueemi99 as the original publisher. QualityMax is not affiliated with or endorsed by Block or Dessa.
Two anonymous voice clusters, plus uncertain spans. Select a segment to hear it and follow its timing. Speaker numbers are model labels, not verified identities.
Select a span above to inspect the model’s assignment.
| Span | Start | End | Model label | Action |
|---|
This explorer shows precomputed analysis. Playback uses the original mixed recording; selecting a voice span does not isolate that voice. This page does not transcribe audio, identify people, or run the Dessa model.
One recording can support specific observations. Broader claims need repeatable tests, known speaker labels, and access to a live system.
Signal-level changes in selected recording windows. Speech boundaries and anonymous voice groups estimated with Silero VAD and WeSpeaker embeddings.
Whether the response fits the request, a cough is audible, or the conversation resumes appropriately. We have not assigned automated verdicts.
Live response latency, interruption handling, reliability across conversations, and human-versus-AI identity are outside this recording analysis.
The original model labels agree with activity on Block’s isolated reference tracks for 96.3% of 89.28 seconds of exclusive active reference audio. This is not overall diarization accuracy or a score for Dessa.
Overlapping reference activity (3.22 s) and silent reference frames are excluded. Anonymous IDs are matched by the best permutation. Alignment offset: +0.55475 s; both comparisons use a fixed −40 dBFS gate and 20 ms frames. Overlapping voices are not separated in the analyzed mono recording.
Selected source-recording windows: 122.38–124.00 s, 128.70–137.63 s, and 158.55–168.24 s. The full evidence export includes raw level measurements, span boundaries, model provenance, and source-file hashes.
Window selection is manual. The earlier local experiment used a proposed +3 dB threshold; this public case study reports observations without turning that threshold into an agent-quality verdict.
Download the evidence JSON ↗An independent research case study exploring how voice behavior could be evaluated. The public page uses precomputed results; it is not a live QualityMax testing feature.