Your recording choice
Updated October 13, 2026. SpeakEse is in private beta. Recording starts after you choose Start or take an eligible recording action from a notification. Recording hours set availability and reminders; they do not turn on the microphone by themselves. Get permission from other people where recording consent is required.
Your voice sample
For new voice enrollment, SpeakEse asks for cloud-processing permission after you finish recording and before uploading your sample. OpenAI checks a short selected sample for audible speech and speaker consistency. AWS holds accepted audio for a 30-minute review window. Only Save my voice activates the reviewed 2–10-second sample for future voice matching. Expired review samples cannot be played or saved, and cleanup removes them; outages can delay deletion. You can replace or delete your voice sample, or turn off cloud-processing permission, in My Voice. Requests already in progress may finish, and existing data stays until deleted or expired.
Normal SpeakEse service
The phone uses Silero and YAMNet to keep likely speech. Approved clips are uploaded privately. The backend uses your My Voice reference with ReDimNet to find possible parts of your speech; cloud speaker separation later confirms which turns belong to you. For new conversations using the continuous processing flow, selected candidate audio is transcribed in small nearby sections while recording continues. A small cloud text check decides language, useful Notes material, and whether a section is too sensitive, mixed, unclear, or likely media such as TV. Only permitted sections continue into automatic Learning or a Notes suggestion.
Personal-topic choices
General Notes suggestions are on by default. Career & feedback, Health & wellbeing, Money & finances, and Legal & official matters each have separate switches for English Learning and Notes suggestions, and those personal-topic switches start off until you choose them. Intimate conversations, passwords or security codes, and another person’s sensitive private information are excluded from automatic Learning and Notes suggestions. Unclear or mixed sections are held instead of treated as safe.
Learning and titles
After speaker confirmation, SpeakEse counts only permitted English words that belong to you. New continuous sessions use a 20-word minimum before final Learning reconciliation. Final reconciliation uses permitted context to improve transcription wording while preserving the wording you actually used, including grammar mistakes, fillers and unfinished speech. It also creates one short neutral conversation title. A conversation can finish with no learning tips.
Notes
A Notes card can appear only for reviewed sections that are suitable for Notes under your choices. The card does not generate a summary. When you tap Generate, SpeakEse can consider the full linked approved conversation while it is still available. Existing Notes-safe transcript and speaker work is reused. Any previously unexamined section is first transcribed and checked for language, privacy category and Notes eligibility; excluded or unclear sections are not used in the final Note. Only remaining safe speaker coverage is processed. Notes can include a summary, decisions, key points and supported action items. SpeakEse does not guess an action owner or deadline when the conversation does not support it.
Source retention
For new conversations using the continuous processing flow, approved cloud audio and full processing transcript/source caches are scheduled for deletion 7 days after the conversation finishes. A service or storage outage can delay cleanup. Temporary working files are removed after processing, with abandoned temporary files cleaned separately. Generated Notes and saved Learning/progress can remain after source audio expires until you delete the related data. Audio playback and new Notes generation stop when the source expires. Older recordings may keep their earlier retention behavior until separately migrated or deleted.
Diagnostic sharing OFF
If Diagnostic sharing is off, SpeakEse still records, identifies your speech, transcribes, creates learning insights, supports playback while source audio is available, and can generate eligible Notes. The phone does not upload rejected YAMNet candidate audio for debugging and does not create a separate cloud diagnostic archive with detailed filter evidence. Minimum service records such as approved product audio, upload receipts, processing state, safe error codes, timings, model/app versions and provider-call accounting can still be retained as required by the product flow.
Diagnostic sharing ON
If you choose to share diagnostics, SpeakEse can temporarily keep detailed filter and processing evidence and upload YAMNet-rejected candidate clips so authorized SpeakEse administrators can investigate recording, filtering, speaker, transcription and learning problems. Approved product audio is not uploaded a second time only for diagnostics. New diagnostic evidence is kept for up to 7 days. Turning sharing off stops new diagnostic collection and requests cleanup. An outage can delay cleanup. Admin can see the choice but cannot enable it for you.
Cloud AI processing
OpenAI processes selected approved-audio excerpts for transcription and speaker separation, selected text for language/privacy/category checks, permitted text for final reconciliation and Learning, and selected conversation material when you request Notes. Asking a cloud model whether text is sensitive means that text has already been sent to that provider for the check. SpeakEse therefore does not claim that sensitive speech is guaranteed to stay on the phone. Amazon Web Services provides private storage, database and account services.
Practice, Admin access and deletion
Spoken practice uses on-device speech recognition; its practice audio is not saved or uploaded. Submitted practice answers are cleared after evaluation or by the pending-answer expiry worker after 15 minutes; outages can delay cleanup. Authorized administrators can inspect retained diagnostic evidence only when sharing allows it. AI debug reports label missing evidence instead of guessing, and preparing a report does not automatically send it to an AI service. You can delete a conversation, remove My Voice, delete Learning data or delete your account. Previously downloaded reports, provider retention and database backups cannot always be recalled immediately. We do not sell information or use it for advertising.
Questions
No security measure guarantees absolute protection. Contact support@speakese.com for privacy questions, a copy of your data or deletion help.