Before you practice
Your voice and data
SpeakMax is a live practice tool without user accounts. Here is what the current product sends, stores, and lets you control.
Updated 30 September 2026
Audio and transcription
Your microphone is used to capture your speech during a room. Cloudflare audio is the first option offered in every voice room. After you explicitly choose it, microphone speech segments (including scenario dictation and Live Group turns) are uploaded through our Worker to Workers AI for transcription. No microphone audio is uploaded to this route before that choice. The audio relay does not store recordings or log audio/transcript content. Cloudflare processes the uploaded audio under its own terms. Cloudflare consent survives reloads of the same room in this tab and is cleared when you leave. A different room requires a new choice. If you choose local Moonshine transcription, the model runs in your browser after its files download. In Chrome, browser speech recognition may process audio through a browser-managed online service. The browser, not SpeakMax, controls that service.
In a live group room, other participants can see committed captions. A live partial caption appears beside the current speaker. Do not share private information you would not want another participant to read.
AI replies and voice output
Finalized text, the current topic or situation, and relevant room history are sent through the SpeakMax Worker to an available AI provider for replies, coaching, or evaluation. A provider may apply its own logging and retention rules. A personal AI key sends these requests to the provider and model you select; if that route fails, SpeakMax may try its hosted providers unless its allowance is reached.
Local Kokoro downloads model files and generates spoken replies in your browser. Browser voice uses a voice supplied by your device or browser. If you explicitly select ElevenLabs, reply text and your key are sent to ElevenLabs to generate audio. In the AI companion, choosing Cloudflare audio selects hosted transcription and voice; reply text is sent through our Worker to Workers AI when spoken replies are enabled. You can choose another voice independently in settings. Human group participants are represented by captions, and optional AI panelists use device voice. Cloudflare audio has a shared daily allowance and a rolling 24-hour guest allowance. The meter shows an estimate and the exact local reset time; Audio options explains the limits. Guest allowances are tied to the signed session, not a person or account. These counters are operational estimates, not an account billing ledger; a new guest token is not the same personal account. The quota ledger stores daily counts per token fingerprint, not audio or transcript content. Captions remain available when speech output is off or unavailable.
What stays in a room
Open Discussion and Focused Practice have separate transcripts inside the current voice room. Room state is stored by a Cloudflare Durable Object so a brief disconnect can be recovered. When the last person leaves a live or solo room, SpeakMax deletes that room's stored state. A voice room can stay recoverable for up to five minutes after an accidental disconnect; using Leave clears it without that wait. SpeakMax does not currently offer an account-wide session library.
Your display name and voice preferences can remain in this browser's local storage. Your signed room token, an optional ElevenLabs key, and some session data stay in this tab's session storage. A personal AI key is kept in tab memory and the active room connection, not saved to local or session storage. A model download may remain in browser cache until you clear site data or the browser evicts it.
Site services and your choices
Cloudflare Turnstile helps protect room entry. Vercel Web Analytics and Speed Insights collect site usage and performance data. SpeakMax also logs operational events, such as AI provider failures and budget limits, to troubleshoot the service. These events are designed without transcript text or API keys.
You can turn off the microphone or spoken replies, use typed input, choose local transcription where available, and leave a room at any time. To remove remembered preferences and downloaded models, clear this site's storage in your browser. Leaving a room does not delete data already processed or retained under an AI provider's or browser speech service's own policy.