ElevenLabs

The five speech models are not a menu of near-identical options; they trade against each other in ways you can feel. Flash v2.5 is the default because it is the one you want most of the time — around 75ms of latency, 32 languages, and a 40,000-character ceiling, which is enough for a long-form narration in a single call. ElevenLabs v3 is the expressive model, the one that carries a performance rather than a read, and its 5,000-character limit reflects that: you send it a scene, not a chapter. Multilingual v2 sits between them at 29 languages and 10,000 characters when fidelity matters more than speed. Turbo v2.5 is deprecated upstream and kept pinnable only so old work still reproduces — Flash v2.5 replaces it at the same price and lower latency. Text to Dialogue is the odd one out and the most useful surprise: send an ordered array of lines with a voice on each, up to 10 distinct voices and 2,000 characters, and you get back a single file where the speakers actually respond to each other, with matched prosody across the cut and per-word timings if you ask for them. Lines take v3 audio tags like [whispers] and [laughs].

Finding the right voice is usually the harder half of the job, so the picker treats it as a real search. The Library tab queries the ElevenLabs Voice Library with facets for language, gender, accent, age, use case and style, and previews play one at a time so you can audition down a column. The Designed tab does the other thing: describe a voice in words — age, accent, texture, pace — and ElevenLabs returns several previews to choose from before you save one permanently to your account. Cloned voices live in the same tab, and once saved, any of these voices behaves identically everywhere else in Genfire. Music works the same way at a larger scale. Music v2 takes a prompt and a length anywhere from three seconds to ten minutes, or a composition plan: chunks with their own text, section markers, per-chunk styles and durations. Ask it to store the song and you get an id back, which lets a later plan regenerate a single range instead of rerolling the whole track. It will also score a video you hand it instead of a prompt. Genfire's Music Video Studio leans on exactly that structure — the section boundaries in the plan are what the cuts get timed to.

The point of putting all of this behind one balance is that audio in Genfire is rarely the finished product. A voice you pick here narrates faceless reels and explainers, becomes an influencer model's standing voice, and gets attached to a video in the editor. The speech you generate is the audio track a lip-sync engine drives a performance from. Voices are reusable as references too: hand a short clip to a video model that conditions on audio and an on-camera character speaks in that voice. Sound effects run from half a second to 30 seconds and can be generated loopable, which is what you want for a room tone or an engine bed rather than a one-shot. Scribe v2 closes the loop from the other end — audio, video or a YouTube URL in, transcript out, in 90+ languages, with a speaker id on every word for up to 32 speakers, keyterm biasing for names your transcript keeps getting wrong, and entity detection or redaction for PII. All of it runs in the browser studio, from the GenBar, as a node in the workflow editor, on the REST API, through the MCP server at mcp.genfire.ai, and from the @genfire/cli — no ElevenLabs account or separate subscription. Genfire is pay-as-you-go: credit packs start at $19, purchased credits last 12 months, optional monthly plans start at $29, and every purchase includes a commercial license.