Transkripsie -- Afrikaans Transcription

Transkripsie uses artificial intelligence to transcribe, organise, and help you understand your recordings. This page fulfils our transparency obligations under Article 50 of the EU AI Act by disclosing when you interact with an AI system and how each feature works.

Risk classification: Transkripsie's AI features are limited-risk systems subject to transparency requirements (Art. 50). They are not classified as high-risk AI systems under Annex III of the EU AI Act. Our transcription and summarisation tools are assistive productivity features — they do not make autonomous decisions about individuals, allocate resources, or perform biometric identification for law-enforcement or access-control purposes. Voice ID is opt-in speaker labelling for your own library, processed on-server without sending audio to a third party.

For how we handle personal data, see our Privacy Policy. For third-party providers that process data on our behalf, see Sub-processors.

AI features

Every AI-powered capability in Transkripsie is listed below. Each entry describes the purpose, when it runs, which provider processes your data, and how you stay in control.

Live transcription

Purpose: Convert spoken audio to text in real time so you can follow a conversation as it happens, search keywords, and export a transcript when the session ends.
When used: When you start a live recording from the dashboard or join a live session. Audio is streamed to our speech-to-text provider during the session.
Providers: Soniox, Inc. — model stt-rt-v4 (real-time streaming speech-to-text).
Human oversight: You control when recording starts and stops. After the session, you can edit any segment, rename speakers, and delete the recording. Live captions are assistive — always verify critical content against the source audio.
Related: Sub-processors · Privacy Policy

Batch transcription

Purpose: Transcribe uploaded audio or video files after upload, including long-form recordings and files imported from other sources.
When used: When you upload a file for transcription or when a meeting-bot recording finishes and is sent for post-processing.
Providers: Soniox, Inc. — batch speech-to-text API.
Human oversight: Transcription runs only on files you upload or recordings you authorise. You can edit the full transcript, correct speaker labels, re-run processing, and permanently delete the recording.
Related: Sub-processors · Privacy Policy

Speaker diarization

Purpose: Identify when different people are speaking and assign speaker labels to transcript segments, so multi-party conversations are readable.
When used: Automatically during live and batch transcription when multiple speakers are detected. Labels appear as "Speaker 1", "Speaker 2", etc. until you assign names.
Providers: Soniox, Inc. — speaker separation bundled with transcription.
Human oversight: You can rename, merge, or split speaker labels in the transcript editor. Diarization is a best-effort assist — verify speaker attribution for any decision that depends on who said what.
Related: Sub-processors · Privacy Policy

Voice ID (speaker identification)

Purpose: Match detected speakers in a recording to voiceprints you have enrolled, so recurring participants can be labelled automatically across recordings.
When used: Only after you affirmatively opt in under Settings → Transcription, on plans that include Voice ID, and only for users outside Illinois (BIPA). Audio embeddings are computed on our server — raw audio for voice matching does not leave the VPS.
Providers: SpeechBrain ECAPA-TDNN — runs on Transkripsie's Hetzner VPS in Germany. No third-party voice-ID provider receives your audio.
Human oversight: Opt-in only; disabled by default. You can turn off speaker identification at any time in Settings (this deletes enrolled voiceprints). You can override any automatic label, enroll or remove voices, and delete recordings containing voiceprints.
Related: Sub-processors · Privacy Policy

Ask AI (single recording)

Purpose: Answer natural-language questions about one transcript — e.g. "What were the action items?" or "Summarise the budget discussion" — using the transcript text as context.
When used: When you open Ask AI on a specific recording and submit a question. Only the transcript text (not raw audio) is sent to the language-model provider.
Providers: OpenRouter, Inc. (primary), with automatic fallback to xAI and Anthropic if the primary provider is unavailable. Model selection is managed by Transkripsie.
Human oversight: You choose which recording to query and what to ask. Responses are generated suggestions — review before sharing or acting on them. You can delete the recording and its transcript at any time.
Related: Sub-processors · Privacy Policy

Ask AI (cross-recording)

Purpose: Search and reason across multiple recordings you own — e.g. "What did we decide about the vendor across last month's calls?" — by retrieving relevant transcript excerpts as context.
When used: When you use the cross-recording Ask AI workspace and select a scope of your recordings. Relevant transcript excerpts are sent to the language-model provider; audio is never sent.
Providers: OpenRouter, Inc. (primary), with automatic fallback to xAI and Anthropic.
Human oversight: Limited to recordings in your account that you can access. You control which conversations are included via your library. Treat answers as drafts; verify against source transcripts before relying on them.
Related: Sub-processors · Privacy Policy

Auto titles

Purpose: Generate a short, descriptive title for a recording based on its transcript so your library is easier to browse.
When used: After transcription completes, if the recording has no title or still carries a default placeholder. You can also trigger re-generation from the recording page.
Providers: OpenRouter, Inc. (primary), with automatic fallback to xAI and Anthropic.
Human oversight: Titles are suggestions only. You can edit or replace the title at any time in the recording editor.
Related: Sub-processors · Privacy Policy

AI overview / summary

Purpose: Produce a structured summary of a recording — key topics, decisions, and action items — using a template you select (or the workspace default).
When used: When you request an AI overview on a recording, or when your workspace template is applied after transcription. Transcript text is sent to the language-model provider.
Providers: OpenRouter, Inc. (primary), with automatic fallback to xAI and Anthropic.
Human oversight: Summaries are assistive drafts. You can edit, regenerate, or delete them. Custom summary templates let you steer tone and focus, but you remain responsible for verifying accuracy.
Related: Sub-processors · Privacy Policy

Translation (to English)

Purpose: Translate non-English transcript segments into English so mixed-language recordings are readable in one language.
When used: When you request translation on a recording with non-English segments. Only text segments are sent for translation; audio is not sent.
Providers: OpenRouter, Inc. (primary), with automatic fallback to xAI and Anthropic.
Human oversight: Machine translation can miss nuance, idioms, and domain terms. Use the original-language transcript as the authoritative source. You can edit translated text and delete translations by deleting the recording.
Related: Sub-processors · Privacy Policy

Meeting bot

Purpose: Join a scheduled video meeting on your behalf (Google Meet, Zoom, Microsoft Teams, or Webex), capture the session, and produce a transcript and recording afterward.
When used: When you connect a calendar, schedule a bot for a meeting, or manually send a bot to a meeting URL. The bot uses platform-native live captions during the call and sends the captured audio for post-meeting transcription.
Providers: Recall.ai, Inc. (meeting bot infrastructure and platform captions during the call); Soniox, Inc. (post-meeting speech-to-text and diarization on the captured audio).
Human oversight: You choose which meetings receive a bot and can cancel before join. Inform other participants as required by your organisation and local law. You can edit the resulting transcript, delete the recording, and disconnect calendar access at any time.
Related: Sub-processors · Privacy Policy

Your rights and controls

Edit transcripts — correct any AI-generated text, speaker labels, or titles before sharing.
Disable Voice ID — turn off speaker identification in Settings; opting out deletes enrolled voiceprints.
Delete recordings — remove audio, transcripts, summaries, and voiceprints from your account (backup copies expire within 30 days).
Export your data — download a ZIP of your profile, transcriptions, and audio from Settings → Security.
No model training on your content — our contracts with Soniox and OpenRouter prohibit using customer audio or transcript text to train their models.

Contact

Questions about our use of AI or this transparency notice: privacy@transkripsie.com

AI Transparency