RESEARCH AUDIO / SOURCE: BLOCK / CHECKED OCTOBER 25, 2026
Dessa AI samples: a practical listening guide
These Dessa AI samples introduce the speech model developed by Block. Listen to complete exchanges to understand the timing and delivery that make a conversation feel responsive. The six recordings below are served from Block’s public website and remain Block research material. They are distinct from the GPT-Live and HeyGen tools available in our independent Dessa AI voice Studio.
You can listen without creating an account or buying credits. Each recording has a short listening prompt; the original project provides synchronized subtitles. The notes below suggest what to examine, rather than presenting a new transcript or claiming that we ran a controlled benchmark. Follow the source link for the original context.
Six Dessa AI samples from the research release
Start with an ordinary exchange, then try one where the subject changes the speaker’s delivery. For these Dessa AI samples, compare the beginning of a reply with the end of the preceding turn. Listening on headphones can help you distinguish the two speakers, especially when speech overlaps. The player loads media only when you request playback.
BLOCK RESEARCH RECORDING
A conversation about the release
Listen for the relationship between laughter, the speaker’s timing, and the next response. This Dessa AI recording is useful for thinking about conversational delivery, rather than testing factual accuracy.
Original recording and synchronized subtitles at BlockBLOCK RESEARCH RECORDING
Working toward a remembered movie
Follow how the conversation develops when a speaker is searching for a title. Notice when a reply begins and whether it helps the exchange continue. A selected example is not a general knowledge benchmark.
Original recording and synchronized subtitles at BlockBLOCK RESEARCH RECORDING
Responding to a surprise
Compare changes in pacing and vocal expression with the topic of the conversation. Listen through the exchange rather than judging Dessa AI from one emotional reaction or a short isolated excerpt.
Original recording and synchronized subtitles at BlockBLOCK RESEARCH RECORDING
Variation in conversational delivery
Listen for shifts in speed and intonation across the recording. Distinguish what you hear in this particular conversation from a claim that any model can consistently produce a requested emotion.
Original recording and synchronized subtitles at BlockBLOCK RESEARCH RECORDING
A personal conversation
Consider how pauses and delivery influence your impression of a response. A voice that sounds attentive does not establish that its advice is correct or appropriate for an important decision.
Original recording and synchronized subtitles at BlockBLOCK RESEARCH RECORDING
Dessa in conversation with GPT-Live-1
Both participants are AI voices in this official recording. Use it to observe turn exchange; it is not a controlled head-to-head test with identical prompts, settings, and independent scoring.
Original recording and synchronized subtitles at BlockHow to evaluate a voice conversation
A pleasant voice is only one part of a useful interaction. Dessa AI samples are more informative when you separate how a reply sounds from what it says. A response can be expressive and still miss a question. Conversely, a plain delivery can carry a precise answer. Consider both qualities before deciding what matters for your own task.
- Timing: notice pauses, overlapping speech, and the point where the next speaker enters. Does the exchange continue smoothly, or does someone have to restart?
- Delivery: listen for changes in pace, emphasis, and intonation. Check whether those changes follow the conversation rather than judging one dramatic moment.
- Relevance: ask whether the reply addresses the preceding turn. Tone alone does not demonstrate understanding, reliable knowledge, or sound advice.
- Recovery: look for moments of uncertainty or interruption. A complete exchange reveals more about recovery than an isolated highlight.
If you take notes on Dessa AI samples, record the clip name, timestamp, and the specific behavior you heard. Keep an observation such as a reply beginning during a pause separate from an interpretation about the underlying architecture. That makes your notes easier to check against the source and compare with another recording.
What Dessa AI samples can and cannot establish
Block describes the public examples as recordings of live conversations and says subtitles were added afterward. These are useful demonstrations of its research model. The released selection does not tell us the rate of successful conversations across all topics, microphones, accents, or environments. We have not measured latency or scored answer quality from these recordings.
The conversation with GPT-Live-1 also has a narrower purpose than a leaderboard. Two AI voices talking to one another do not create an identical test for both models. Settings, conversation roles, and evaluation criteria matter. Read our Dessa vs GPT-Live comparison for a framework that keeps published demonstrations and product choices distinct.
These Dessa AI samples also do not establish access rights. A public recording is not a model download, an API key, or permission to join a research program. The official access guide explains what we found on Block’s project page and where to verify changes.
From listening to your own voice task
After hearing the Dessa AI samples, decide whether you need a spontaneous exchange or a repeatable reading of a script. For conversation practice, start with a focused topic and review your transcript afterward. For narration, prepare a short script with punctuation and names checked before generating a full version.
Our Studio offers these two workflows through different providers. It does not reproduce Block’s research model. Review the Studio plans and credit rules before a paid operation, and check the provider named in the tool. Research sample playback is separate from your account balance.
Open the independent voice Studio