Double-click start.bat, or from a terminal:

start.bat

Python install: download from python.org and ensure "Add python.exe to PATH" is enabled — otherwise start.bat cannot locate the interpreter.

chmod +x start.sh
./start.sh

Python install: the Python shipped with macOS is typically outdated. Install a current version from python.org or via Homebrew (brew install python@3.13).

chmod +x start.sh
./start.sh

start.sh detects missing python3-venv / python3-pip and offers to install them via apt / dnf / pacman.

For a native desktop window (otherwise opens in your default browser):

sudo apt install gir1.2-webkit2-4.1 python3-gi    # Debian/Ubuntu

Limitations:

• PDF export unavailable on Linux (relies on Microsoft Word) — use DOCX instead.
• Without a system keyring, API keys live only for the running session. Either start a keyring daemon, or rely on the bundled keyrings.alt file fallback.

For active development, use dev.bat (Windows) or ./dev.sh (Linux/macOS) — runs the app under shiny run --reload, auto-restarting on .py saves.

The first screen on launch. A workflow strip (Audio → Transcript → Analysis → Feedback → Export) shows how far you are; entry tiles jump to a T-SEDA analysis (loads the built-in codebook with its presets — just add a transcript), transcription, analysis, the session history, or a no-key demo. A configuration line reports the active provider/model, and a one-time data-protection acknowledgment (pick explicit consent, fictive test data, or only my own utterances) gates LLM calls until confirmed.

Document Input panel:

• Transcript (required) — must follow the noScribe format. The interactive multi-stage converter handles transcripts from other tools (e.g. aTrain) — speaker-label detection, timestamp stripping, bracket-annotation review, per-speaker mapping, side-by-side preview.
• Codebook (required for qualitative analysis) — see the example codebook. Codes apply to all speakers (teacher and students). Alternatively, load the built-in T-SEDA template with one click (official T-SEDA codes for dialogic classroom talk, DE/EN; T-SEDA Collective 2023, University of Cambridge, CC BY) — it also presets coding of teacher and student turns (context-aware) plus multi-coding with confidence.
• Teacher name (optional) — if present in the transcript, enables teacher-specific metrics.
• Group identifier and metadata — used for report labelling.

Configure the LLM and click Analyze in the tab's LLM configuration card; the app switches to Results on completion.

Cost estimate. The figure in the LLM-configuration card is a lower bound — transcript+codebook length × input price × ~4 for output. A cumulative cost tracker (in Options) sums spend across all your analyses.

Turn an audio recording into a transcript entirely on your machine — the audio never leaves your computer. Powered by the standalone open-source engine noScribe (Whisper + pyannote), GPL-3.0, invoked only as a separate subprocess and installed on demand (~3 GB, one time, Windows).

Exposes the full set of noScribe options: audio in, output filename, start/stop range, language, model (fast / precise), speaker count (pre-filled from the group size), mark-pause, overlapping speech, disfluencies, timestamps. An in-app waveform editor lets you drag handles for start and end of the segment — no external tool needed. A live progress display (step, percentage, elapsed time) shows what's happening, and an editable transcript field lets you fix speaker labels or spelling before the result is handed straight into the Analysis tab or saved as a .txt file.

Best suited to 10–15 minute small-group recordings (CPU transcription ≈ 1.5× realtime with the fast model). See the in-app Info / License tab and NOTICE for the licensing and privacy details.

Generate research-grounded, formative feedback for the teacher after an analysis has run. The Feedback tab takes the coded turns, the codebook definitions, and the quantitative metrics from the same session and produces a structured prose report along three axes — Stärken (strengths) · Entwicklungsfelder (growth areas) · Konkrete Umsetzungshinweise (concrete suggestions) — with a short reference list grounded in dialogic-teaching literature (T-SEDA, IRE/IRF, accountable talk, productive disciplinary engagement).

The generated text is fully editable in-place — you can tighten, rephrase, or strip sections before exporting to Word (.docx) or PDF. A live cost estimate and the cumulative cost tracker (in Options) apply here too.

You can also upload a corrected report (DOCX / XLSX / CSV / HTML) so the feedback rests on the codings you reviewed in Word or Excel rather than the model's raw output — it doubles as a standalone entry point in a fresh session (codes and metrics come out of the document, and the codebook is rebuilt from the report legend if present).

It is an aid, not a verdict — clearly framed as formative scaffolding for self-reflection, not summative assessment.

Generate a print-ready GDPR Art. 13 consent declaration for the training context — where a trainer team works with teachers and each teacher consents to processing their own recording. A pre-filled form (left) renders a live document preview (right) that you export to Word (.docx) or PDF.

The declaration reflects the real data flow: local transcription (audio stays on device) vs. the configured LLM as recipient. A cloud/local toggle drives the third-country transfer paragraph and a separate consent checkbox; missing mandatory fields surface as red !!! … !!! markers. The wording is adapted from the CC0-licensed Consent-Gen-RDMO (TU Dortmund).

It is an aid, not legal advice — the disclaimer is shown in the form and the document footer; have it reviewed by your data protection officer before use. See NOTICE.

Split into quantitative and qualitative sections.

Quantitative (deterministic): participation metrics, conversation shares (absolute + relative), per-speaker turn stats (count / mean / median), three-segment over-time view.

Qualitative (LLM-coded): per-speaker coding (every turn carries a Sprecher label), code distribution plot stacked by speaker group (teacher vs. students, with the matching code × speaker-group table in the report), coded-impulse table, over-time code distribution, Markov-style code-transition heatmap, and an auto-generated methods paragraph for paper manuscripts (copy-to-clipboard, EN/DE). Every coding carries the model's confidence (e.g. EN (92 %)) — in single-coding on the one code, and with multi-coding enabled across the top 2 candidate codes in dedicated columns (matching T-SEDA's 0–2-codes-per-turn rule), ranked by confidence; uncertain candidates stay visible, the human judges. The DOCX/HTML report shades the certainty edges of each code cell (≥ 90 % and < 50 %) using the same anchors the prompt calibrates against. A second review pass automatically re-submits uncoded turns to the LLM for a careful re-check, and every cell stays hand-editable. All turns of the conversation appear in the table and reports, coded or not.

• API configuration — keys for LocalMind, OpenAI, Anthropic, Mistral, DeepSeek, plus any number of your own custom OpenAI-compatible endpoints (add/rename/delete them in Options — each with a name, base URL and its own key, e.g. a self-hosted vLLM server and an institutional gateway side by side). Keys live in the OS keyring (Keychain / Credential Manager / SecretService); names and base URLs persist in your local config.
• Models for LLM Selection — edit the registry (add/remove models, set per-million-token prices); changes propagate to the Analysis-tab model picker in real time. A Load models from provider button refreshes the selected provider's list straight from its live catalogue (needs a saved key); embedding/audio/image models are filtered out and prices you already set are kept.
• Custom Prompts — modify the system + user prompts used for qualitative coding; defaults restorable any time.
• Cost tracker — cumulative spend across all analyses, per provider.
• Test the app (gold-standard self-test) — runs a known fixture and shows expected vs. actual to build trust before you analyse real data.
• Additional Options — defaults for teacher name, group ID, class size, advanced toggles like streaming.