How We Curated 279 AI Voices for DIALOGUE: Behind the Voice Library

<p>When DIALOGUE launched, we offered 30 built-in AI voices. That was fine for an early product — enough variety to cover a few languages and content types. But as our users started producing podcasts in 7 languages, across dozens of formats, from company newsletters to course modules, 30 voices started feeling like a creative ceiling.</p> <p>So we did something unusual. Instead of licensing a handful of premium voices from a single provider, we paged through the entire ElevenLabs shared voice library — every page, every voice — and curated a catalog of 279.</p> <h2>Why the ElevenLabs Shared Voice Library?</h2> <p>The shared voice library is a community resource. These are <strong>real voices contributed and used by creators worldwide</strong> — not synthetic voices generated from a handful of base models. Each one has a voice_id that maps to an actual ElevenLabs voice you can look up. That transparency matters to us: when you select a voice in DIALOGUE, you're choosing a real voice that exists in the ecosystem, not a black-box label.</p> <p>It also means the voices have been battle-tested. If a voice has been used thousands of times across the ElevenLabs platform, that's a strong signal it sounds good in practice, not just in a 10-second preview.</p> <h2>How We Auditioned Every Voice</h2> <p>The process was methodical, not algorithmic. We paged through the shared voice library systematically, testing each voice against four criteria:</p> <p><strong>Natural sound quality.</strong> Does it sound like a person talking, or like a machine reading? We rejected voices with robotic artifacts, uneven pacing, or synthetic timbre.</p> <p><strong>Expressive range.</strong> Can the voice handle questions, statements, excitement, and calm explanation without falling flat? A voice that only does one emotion well isn't versatile enough for podcasting.</p> <p><strong>Clarity.</strong> Can you follow complex sentences without rewinding? This is especially important for educational and business content where dense information is common.</p> <p><strong>Language authenticity.</strong> For non-English voices, we prioritized native speakers. A Japanese voice that sounds like it was built from an English TTS model with Japanese phonemes glued on top is immediately noticeable to a native listener — and it breaks trust.</p> <h2>The Descriptive Label System</h2> <p>One of the messiest parts of voice libraries is free-text descriptions. "A warm but professional voice with occasional playfulness" — what does that actually mean? Different people hear different things.</p> <p>We replaced that ambiguity with <strong>clean descriptive labels</strong> — chill, casual, calm, confident, deep, energetic, warm, authoritative, and others. Each voice gets exactly one primary label. This isn't a reduction; it's a design choice. When you're scanning 279 voices for a tech news podcast, filtering by "energetic" gets you to the right candidates faster than reading paragraphs of prose.</p> <h2>Usage-Based Ranking</h2> <p>Voice order in our catalog isn't arbitrary. We rank voices by <strong>real-world popularity</strong> using ElevenLabs's usage data. The voices that creators worldwide reach for most often surface at the top. This means the catalog self-improves over time — as usage patterns shift, the ranking reflects what's actually working in practice.</p> <h2>Accent Curation</h2> <p>AI voice catalogs often default to a single "standard" dialect per language. We pushed for more. Our Japanese voices include Kansai dialect alongside standard Japanese. Vietnamese voices cover both Southern and Northern accents. This isn't just about representation — it's about <strong>audience trust</strong>. A listener in Osaka hears the difference between a Kansai speaker and a Tokyo-standard voice, and that difference affects how they receive your content.</p> <h2>Gender Balance</h2> <p>For every language in the catalog, we aimed for roughly equal male and female voices. This isn't a checkbox exercise. Two-host podcasts work best when the voices have distinct but complementary characteristics, and gender is one axis of that contrast. A balanced catalog means you can pair voices for dynamic conversations without compromise.</p> <h2>Why 279, Not 500?</h2> <p>We could have included more. The ElevenLabs shared voice library has thousands of entries. But we set a quality bar and stuck to it: <strong>every voice was personally auditioned and rated</strong>. Adding more voices below that bar would inflate the number without improving the experience. 279 was the natural ceiling where quality and quantity intersected.</p> <h2>Transparency</h2> <p>Every voice_id in our catalog maps to a real ElevenLabs voice. No synthetic labels, no opaque internal IDs. You can take any voice_id from DIALOGUE and look it up in the ElevenLabs ecosystem. We think creators deserve to know exactly what voices they're using.</p> <p>Curating 279 voices wasn't the fastest path — it took weeks of listening, comparing, and rating. But the result is a catalog where every voice earns its place, sorted by what actually works, in the dialects people actually speak. That's the library we wanted as creators ourselves.</p>