README.md

---
title: GTROX
emoji: 🎙️
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
pinned: true
tags:
- track:backyard
- sponsor:modal
---

# GTROX Speech Coach

GTROX analyzes recorded speech and returns a transcript, measurable delivery
KPIs, and personalized coaching.

## Pipeline

1. The Gradio app accepts an uploaded file or microphone recording.
2. NVIDIA Parakeet runs on Modal and produces the transcript.
3. Local deterministic analysis calculates WPM, filler density, repetitions,
   and a fluency score.
4. Qwen2.5-1.5B-Instruct runs on a Modal T4 and generates grounded coaching.
5. KPI-based coaching is returned automatically if the AI coach is unavailable.

The public transcription service limits uploads to 25 MB and five minutes.

## Project Structure

```text
app.py                  Gradio UI
gtrox/config.py         Endpoint configuration
gtrox/clients.py        Modal HTTP clients
gtrox/metrics.py        Deterministic KPI calculations
gtrox/pipeline.py       End-to-end orchestration
services/transcribe.py  Modal Parakeet service
services/coach.py       Modal Qwen coaching service
tests/test_metrics.py   KPI unit tests
```

## Deploy Modal Services

```bash
modal deploy -m services.transcribe
modal deploy -m services.coach
```

The endpoint defaults are stored in `gtrox/config.py` and can be overridden
with the `TRANSCRIBE_URL` and `COACH_URL` environment variables.

## Current Limitations

Version 1 does not claim to measure pauses, tone, emotion, or pronunciation.
Those features require a reliable timestamp or voice-activity detection layer.