The latest articles on DEV Community by ivyjsu (@ivyjsu). https://dev.to/ivyjsu https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4046036%2F99a3c25f-0999-4ad3-bb21-7daa38dd5328.png https://dev.to/ivyjsu en ivyjsu Fri, 24 Jul 2026 19:20:55 +0000 https://dev.to/ivyjsu/making-two-tts-voices-sound-like-an-actual-conversation-3hn7 https://dev.to/ivyjsu/making-two-tts-voices-sound-like-an-actual-conversation-3hn7 <p>Rendering a document as a two-host podcast is not just alternating two TTS voices line by line. Do that and you get something that sounds like two people reading unrelated scripts in the same room.</p> <h2> The problem is prosody, not voices </h2> <p>Human conversation carries information in timing. A reply that arrives instantly reads as agreement. A short gap reads as consideration. Speakers overlap slightly at turn boundaries. Pitch tends to fall at the end of a statement and rise before a handoff.</p> <p>Concatenated TTS has none of this. Each utterance is synthesised in isolation with neutral prosody and identical gaps, and the result sits in an uncanny valley — clearly speech, clearly not conversation.</p> <h2> Things that measurably help </h2> <p><strong>Vary inter-turn gaps by turn type.</strong> Not one fixed pause:</p> <ul> <li>Agreement or continuation: short, 150-250ms</li> <li>Topic shift: longer, 400-600ms</li> <li>Answer to a question: short, the responder was already primed</li> <li>After a complex explanation: longer, it reads as processing time</li> </ul> <p>A lookup keyed on turn type gets you most of the way, and it is cheap to implement.</p> <p><strong>Write handoffs into the script.</strong> The transition matters more than the voice. "Right — so what happens when the file is large?" carries turn-taking structure that "The file size limit is 10MB" does not. This is a scripting problem before it is an audio problem.</p> <p><strong>Do not centre both voices.</strong> A few degrees of stereo separation makes two speakers legible to the ear. Hard-panning is disorienting on headphones; subtle offset is not consciously noticed but does the work.</p> <h2> Chunking is where quality actually dies </h2> <p>Most TTS APIs have input length limits, so long documents get split. Naive splitting at the limit cuts mid-sentence, and the resulting prosody discontinuity is jarring — the pitch contour restarts mid-thought.</p> <p>Split on <strong>semantic boundaries</strong> — paragraph, then sentence — and never mid-clause. If a single sentence exceeds the limit, split at a comma and accept a small artefact rather than a mid-word cut.</p> <h2> Multilingual complicates the pause table </h2> <p>Turn-taking norms are not universal. Japanese conversation tolerates longer pauses than English; treating a 600ms gap as "awkward" is an English-specific assumption. If you support multiple languages, the pause table needs to be per-language, not global.</p> <p>I build this into <a href="https://getduocast.com" rel="noopener noreferrer">DuoCast</a> — paste text, a PDF, or a URL, get a two-host podcast with the conversational layer handled.</p> <h2> Caveat </h2> <p>None of this makes synthetic conversation indistinguishable from a real recording. It moves it from "obviously robotic" to "pleasant enough to listen to on a commute," which is a lower but more achievable bar.</p> ai audio python nlp