Upload 5 files

.gitattributes

@@ -1,4 +1,6 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
+* text=auto eol=lf
+
+*.7z filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,6 @@
+.venv/
+__pycache__/
+*.py[cod]
+.env
+analysis_output.txt
+*.wav

README.md

@@ -1,42 +1,54 @@
 ---
 title: GTROX
-emoji: 💬
+emoji: 🎙️
 colorFrom: yellow
 colorTo: purple
 sdk: gradio
 sdk_version: 6.5.1
 app_file: app.py
 pinned: true
-hf_oauth: true
-hf_oauth_scopes:
-- inference-api
 ---
 
-# 🎙️ GTROX: Speech Insight Coach
+# GTROX Speech Coach
 
-**GTROX** is an AI-powered speech analysis tool designed to help you improve your fluency, pacing, and clarity. By leveraging advanced speech-to-text models and compact, high-performance Language Models (SLMs), it provides actionable, real-time coaching feedback based on your audio input.
+GTROX analyzes recorded speech and returns a transcript, measurable delivery
+KPIs, and personalized coaching.
 
-## ✨ Key Features
+## Pipeline
 
-* **Fluency Tracking:** Instantly calculate your Words Per Minute (WPM) to master your speaking pace.
-* **Silence Analysis:** Identify how much of your speech consists of pauses and filler time.
-* **AI Coaching:** Get personalized, constructive tips generated by a specialized Language Model to improve your delivery.
-* **Robust Pipeline:** Engineered for high performance, utilizing efficient audio processing without heavy external dependencies.
+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.
 
-## 🛠️ How It Works
+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
+```
 
-1. **Input:** The system processes your uploaded `.wav` file or live microphone recording.
-2. **Analysis:** It calculates technical metrics like WPM and silence ratios using numerical analysis.
-3. **Transcription:** Uses `distil-whisper` for highly accurate, low-latency speech-to-text.
-4. **Feedback:** The `SmolLM2` language model analyzes your performance and suggests improvements.
+## Deploy Modal Services
 
-## 🚀 Deployment
+```bash
+modal deploy -m services.transcribe
+modal deploy -m services.coach
+```
 
-This app is built with **Gradio** and hosted on **Hugging Face Spaces**. It is optimized for low-latency inference, making it suitable for quick fluency checks on the go.
+The endpoint defaults are stored in `gtrox/config.py` and can be overridden
+with the `TRANSCRIBE_URL` and `COACH_URL` environment variables.
 
-## 📦 Requirements
-
-To run this application locally, ensure you have Python 3.10+ and the necessary dependencies installed:
+## 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.

app.py

@@ -1,82 +1,34 @@
 import gradio as gr
-import librosa
-import numpy as np
-import torch
-import soundfile as sf
-from transformers import pipeline
-import os
 
-# 1. Model Initialization
-# We set to 'cpu' as default for broader compatibility, but it will detect 'cuda' if available.
-device = "cuda" if torch.cuda.is_available() else "cpu"
+from gtrox.pipeline import analyze_audio
 
-# Transcriber: Fast, efficient speech-to-text
-transcriber = pipeline("automatic-speech-recognition", 
-                       model="distil-whisper/distil-small.en", 
-                       device=device,
-                       return_timestamps=True
-                       )
 
-# Coach: Small Language Model for feedback
-coach = pipeline("text-generation", 
-                 model="HuggingFaceTB/SmolLM2-1.7B-Instruct", 
-                 device=device)
-
-def analyze_speech(audio_path):
-    if audio_path is None:
-        return "Please upload an audio file."
-
-    # 1. Load Data
-    try:
-        data, sr = sf.read(audio_path)
-        if len(data.shape) > 1: data = data.mean(axis=1)
-        if sr != 16000:
-            data = librosa.resample(data, orig_sr=sr, target_sr=16000)
-    except Exception as e:
-        return f"Error loading audio: {str(e)}"
-
-    # 2. Transcription (Using optimized chunking)
-    # This prevents the "long-form" 30s crash
-    try:
-        # Pass the array and sample rate directly
-        result = transcriber({"raw": data, "sampling_rate": 16000})
-        text = result["text"]
-    except Exception as e:
-        return f"Transcription error: {str(e)}"
-
-    # 3. Metrics
-    duration = len(data) / 16000
-    wpm = (len(text.split()) / (duration / 60)) if duration > 0 else 0
-    silence_ratio = np.sum(np.abs(data) < 0.01) / len(data)
-
-    # 4. Feedback
-    prompt = (
-        f"You are a professional speech coach. The user spoke at {wpm:.1f} WPM with {silence_ratio:.1%} silence. "
-        f"Transcript: '{text}'. "
-        "Provide exactly 2 short, constructive tips to improve their delivery, focusing on pacing and clarity."
+with gr.Blocks(title="GTROX Speech Coach") as demo:
+    gr.Markdown("# GTROX Speech Coach")
+    gr.Markdown(
+        "GPU transcription by NVIDIA Parakeet, delivery KPIs, and personalized "
+        "coaching by Qwen2.5."
     )
-    try:
-        # Simplified generation call
-        feedback = coach(prompt, max_new_tokens=100)[0]['generated_text']
-    except Exception as e:
-        feedback = "Could not generate feedback."
 
-    return f"📝 Transcription: {text}\n\n📊 Metrics: {wpm:.1f} WPM, Silence: {silence_ratio:.1%}\n\n💡 Coach Feedback:\n{feedback}"
+    audio_input = gr.Audio(
+        type="filepath",
+        label="Record or Upload Audio",
+        sources=["microphone", "upload"],
+    )
+    analyze_btn = gr.Button("Analyze Speech", variant="primary")
+    transcript_box = gr.Textbox(label="Transcript", lines=8)
+    metrics_table = gr.Dataframe(
+        headers=["Metric", "Value"],
+        interactive=False,
+    )
+    feedback_box = gr.Markdown(label="Coach Feedback")
 
+    analyze_btn.click(
+        analyze_audio,
+        inputs=audio_input,
+        outputs=[transcript_box, metrics_table, feedback_box],
+    )
 
-# 2. UI Layout
-with gr.Blocks() as demo:
-    gr.Markdown("# 🎙️ ImproveTalk: Speech Insight Coach")
-    gr.Markdown("Upload your `.wav` file to get instant feedback on your fluency.")
-    
-    with gr.Row():
-        # audio_input = gr.Audio(type="filepath", label="Upload WAV Audio")
-        audio_input = gr.Audio(type="filepath", label="Record or Upload Audio",sources=["microphone", "upload"])
-        analyze_btn = gr.Button("Analyze My Speech", variant="primary")
-    
-    results = gr.Textbox(label="Analysis Results", lines=10)
-    
-    analyze_btn.click(analyze_speech, inputs=audio_input, outputs=results)
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()

requirements.txt

@@ -1,16 +1,3 @@
-# UI Framework
 gradio>=5.0.0
-
-# Core AI/ML Libraries
-torch>=2.0.0
-transformers>=4.40.0
-accelerate>=0.30.0
-
-# Audio Processing
-librosa>=0.10.0
+requests>=2.32.0
 soundfile>=0.12.0
-numpy>=1.26.0
-
-# Inference Optimization (Crucial for Small Models)
-bitsandbytes>=0.43.0
-