Voice, Speech & Realtime AI

Block built a voice model trained entirely on human conversation — and it is the clearest signal yet of what interview audio is worth

Source: The Tribune · Oct 24, 2026

Block has introduced a speech model it calls Dessa, and the headline is not the benchmark numbers — it is the training recipe. The company says Dessa was built from scratch on human conversation alone: no text corpus underneath it, no pretrained foundation model to fine-tune, just recorded speech. It is a full-duplex, speech-to-speech system, meaning it listens and responds in audio directly rather than transcribing input to text, generating a text reply, and reading it back. The stated goal is to keep everything the text round-trip throws away: the emotion in a voice, the sarcasm, the laughter, the pauses where someone is thinking, the way real people interrupt and correct themselves.

Whatever one makes of the product, the design choice is the interesting part. For years the default assumption was that the model was the hard part and data was a commodity you scraped. A model trained only on real human conversation inverts that: the architecture is replicable, but the input is not. You cannot synthesize your way to the texture of genuine spoken dialogue, and you cannot scrape it cleanly at the scale or quality a speech model needs. The constraint moves from compute to conversation.

That is exactly where interview audio becomes valuable. A structured interview is one of the few settings that naturally produces long-form, real, technical speech: a person reasoning out loud, being pushed with follow-up questions, backtracking, explaining, under mild pressure, over many minutes. It is not scripted, it is not a single clean utterance, and it comes paired with context — the question asked, the domain, the outcome. Captured with consent and stripped of identifying detail, it maps almost directly onto what a speech-to-speech team needs for both training and evaluation.

The broader signal is that the most ambitious voice work is moving toward learning from authentic human interaction rather than text proxies. As it does, the advantage shifts to whoever can supply consented, richly varied, real-conversation audio with documented provenance. An interview corpus is built from precisely that shape of data — which is why a model trained on nothing but human conversation is less a curiosity than a preview of what the market will be buying.

Key Points

• Block unveiled Dessa, a full-duplex speech-to-speech model it says was trained from scratch on nothing but human conversation — no text corpus and no pretrained base model
• Processing audio directly, rather than through a text round-trip, is meant to preserve the parts of speech that transcription discards: emotion, sarcasm, laughter, hesitation and cadence
• A model that learns only from real spoken interaction makes the training input, not the architecture, the scarce resource
• Consented, anonymized interview recordings are one of the few naturally occurring sources of long-form, real, technical conversation paired with context