πŸ“‹ An PrivilegeScribe schicken β€” Fragebogen ausfΓΌllen
PrivilegeScribe Β· Test Run Walkthrough

Live Hearing β†’ Annotated Noise β†’ Legal Report

Every step of the combined test, with comments and outcomes
Run date 12 Oct 2026 Β· fictitious matter Schmidt v Alpine Bank Β· all local, no cloud Β· model qwen3-coder:30b
πŸŽ™ Live recorder :5003β–Ά πŸ“ annotated noise/β–Ά πŸ“‚ INPUT FILES :5021β–Ά πŸ€– qwen3-coderβ–Ά πŸ“„ Legal report

This page replays the test run of the new live-hearing pipeline: a hearing is recorded with noise annotation, the annotated JSON lands in annotated noise/, appears (newest first) under πŸ“‚ Input Files in the PrivilegeScribe prototype, and is turned into a structured report in EN, FR, DE, ES and IT.

Press β–Ά Play for the animated run, or Show all. The hearing was synthesized with text-to-speech voices; a real microphone still needs one manual check.

Preconditions
Ollama, model, free ports
Terminal
Before anything, the model must answer. Earlier, a model instance already loaded in Ollama returned empty text ("Compute error").
Model installed and answering, both ports free.
Record the hearing
Live hearing for annotated noise Β· 0 Start Live Hearing.command Β· http://127.0.0.1:5003
PrivilegeScribe β€” Live Hearing
Matter ID: Schmidt_v_Alpine   ● recording 00:00
Judge Β· claimant counsel Β· noise Β· respondent counsel Β· judge β€” with a 3.8 s pause
The recording is a fictitious hearing, spoken by three synthetic voices, with a pause and noise bursts added. Audio is transcribed locally with Whisper; nothing leaves the Mac.
Recording, noise detection and Stop work. The annotated JSON is saved to annotated noise/ with 15 segments, timestamps, and audio quality markers.
The annotated transcript is saved
annotated noise/liveHearing_Schmidt_v_Alpine_20261012_215357.annotations.json
annotated noise/ (JSON)
Each segment carries a kind (speech, noise, silence), start time, text and level in dB. Noise and pauses become explicit markers, so the report can say what was lost and when.
JSON saved and opened automatically. 15 segments, 33.6 s, language detected as English. Whisper misheard "Schmidt" as "Schmitt" and "Counsel" as "Council", which is why a human review stays necessary.
The file appears under πŸ“‚ Input Files, newest first
prototype/0 Start PrivilegeScribe.command Β· http://127.0.0.1:5021
PrivilegeScribe β€” πŸ“‚ Input Files
πŸ“„ liveHearing_Meyer… demo_hearing_Meyer_v_Northwind.json17.5 KB
The prototype copies every new JSON from annotated noise/ into its INPUT folder every 3 s and sorts the list by time, newest on top. The list also refreshes itself every 5 s.
Order verified: Schmidt_v_Alpine (Meyer demo kept as archive). The "file not visible" problem earlier was a stopped server, plus the old alphabetical sort.
Prompt, language and matter settings
Prompt: privilegeScribe_text_20261012_liveHearingNoise
PrivilegeScribe β€” Output Language & Register
LANGUAGE (new order)
EN Β· EnglishFR Β· FrenchDE Β· GermanES Β· SpanishIT Β· Italian
MATTER
Schmidt v Alpine Bank Β· Client Schmidt AG Β· Germany Β· Privileged & Confidential
As requested, English now comes first and is the default, followed by French, German, Spanish and Italian (Spanish is new). The matter fields are now passed to the model.
Five language buttons in the right order, EN preselected, new prompt listed among the 12 prompts.
Generate the report (English)
Local qwen3-coder:30b Β· about 1–3 minutes
Generated report β€” EN
All 10 sections are filled. Every claim carries a [mm:ss] reference back to the recording. The three noise events (red) are listed with the words before and after them, and the privilege objection is flagged (yellow) with a recommended handling.
Report complete and faithful. Court, judge and case number show "not stated" instead of invented values.
Multilingual output: EN, FR, DE, ES, IT
Same hearing, five output languages
First attempt vs. after fix
LanguageFirst runAfter fix
ENEnglishEnglish, pause now listed
FREnglish (setting ignored)"DΓ‰TAILS DE L'AUDIENCE" …
ESEnglish (setting ignored)"DATOS DE LA AUDIENCIA" …
ITEnglish (setting ignored)"DETTAGLI DEL PROCESSO" …
DE
The server's language-override function existed but was never called, so the model followed the English prompt. I now append the override, a translate-the-headings instruction and a "list every [PAUSE]" rule to the request.
FR, ES and IT reports now come out in the chosen language, with [mm:ss] and [NOISE]/[PAUSE] markers unchanged. The pause that used to be omitted is now listed.
DE still keeps its English section headings. A DE-specific heading list in the prompt would fix this.
Robustness: empty answers from Ollama
Observed 3 times during the run
Terminal
Sometimes a loaded model returns nothing, so the report is empty without an error message. Unloading the model and retrying fixes it every time.
Workaround known. An automatic unload-and-retry in the server is recommended.