Every voice was Kore.

This weekend we shipped two voice features. Audio Studio now speaks with ElevenLabs: v4, v4 Turbo and Flash v2.5, with all 21 premade voices. Voice input now transcribes with ElevenLabs Scribe v2, at $0.11 per audio hour instead of the $0.36 we were paying for Whisper.

Then we did something we should have done earlier: we pressed every voice preview button in Audio Studio, all 205 of them, and wrote down what came back. Some of it was embarrassing. Here is the feature, then the findings.

1. Three ElevenLabs models, priced per character

ElevenLabs bills text-to-speech by input character, so we do too. You pay the provider's list price for the characters you send, with no markup, and the picker shows a live estimate before you press Generate.

We checked the arithmetic against the provider rather than trusting the rate card: in testing, our per-character charge matched the generation cost OpenRouter reported for each call. Because the price is known before the call, we check your balance against the exact amount first. If it does not cover the request, you get an “add funds” message and nothing is sent upstream.

A few details that took longer than expected:

• The style box is gone for ElevenLabs. Gemini TTS takes a style instruction such as “speak warmly”. ElevenLabs does not; it reads the instruction out loud as part of the script. For ElevenLabs models the picker replaces the style box with a hint: on v4, put tags like [whispering] in the text itself.
• Long scripts take real time. v4 voices generate at roughly 60 characters per second, so a 10,000-character script runs close to three minutes. Those requests get a 300-second timeout instead of the dispatcher default, and anything over the model's limit gets a clear 413 with the limit and your count instead of a provider error.
• One audio format. ElevenLabs returns raw PCM; we wrap it as 24 kHz WAV, the same as our Gemini path, so Audio Studio's join for long multi-part scripts works unchanged.
• No training on your text. Requests route to ElevenLabs with provider data collection denied.

ElevenLabs sits after the Gemini models in the catalog. The default model is unchanged.

2. Voice input at under a third of the cost

The microphone button in chat uses your browser's built-in speech recognition where it can. That path never touches our servers and stays free for everyone. For browsers without it, we transcribe on the server, and that server path now uses ElevenLabs Scribe v2 at $0.11 per audio hour, down from $0.36 for OpenAI's whisper-1. If Scribe fails, the request falls back to OpenAI automatically, so a provider hiccup costs a little more rather than losing what you said.

While wiring this up we found that the server transcription endpoint had no sign-in check. Anyone who could reach it could transcribe audio on our bill. It is closed now: server transcription requires a Pro account, and signed-out requests get a 402.

3. What 205 preview buttons told us

Audio Studio lets you preview any voice on any model before spending money on a full script. We had tested previews the way most people do: a handful of clicks on the models we use. This time we scripted a sweep of every model and voice combination. Three findings stood out.

Every Gemini voice was Kore. The voice picker showed the same 13 OpenAI voice names for every model. Pick “alloy” on a Gemini model and we sent “alloy”; Gemini did not recognize it and quietly used its default voice, Kore. No error, just the wrong voice, every time. The picker now shows the voices each model actually offers, so Gemini lists Gemini voices and ElevenLabs lists its 21.

A voice introducing itself is apparently suspicious. Gemini 3.1 Flash TTS failed the preview on all 30 of its voices. The preview sentence was “My name is Kore and this is my voice”, and Google's content moderation blocked a synthetic voice claiming a name. A neutral sentence passes on every voice. Worse, preview failures had been going to the browser console, so the button simply did nothing. Failures now show in the error banner, and a moderation block returns a plain-language 400 content_blocked instead of a raw 502.

Some “voices” were songs. Lyria, Google's music model, outputs audio, so our catalog filter listed it in Audio Studio. As a voice preview it sang 15 to 90 seconds of music at $0.04 to $0.08 a time, and 7 of 13 attempts returned nothing at all. Lyria is out of the Audio Studio catalog, the speech endpoint now refuses music models with a wrong_studio error, and Music Studio is unchanged.

None of these were exotic. They were the kind of bug that survives because each one fails quietly. The lesson is less about any single fix than about the method: press every button, record what comes back. Our smoke tests now also check that no music model appears in the speech catalog.

4. Studios are now a Pro feature

Image, Audio, Music and Video Studio, and server-side voice transcription, now require a Pro account. On Nyquest, Pro is not a subscription: your first wallet top-up makes your account Pro, and each generation is still paid from your balance at the price shown. Non-Pro accounts get a clear upgrade message before any provider call is made. If you bring your own key for image or audio, the check does not apply.

Audio Studio is in the Nyquest workspace. Pick an ElevenLabs model, press a preview, and it will speak in the voice you chose.

← All engineering posts