🗣️ Briskk Speech-to-Text Assignment

📌 Introduction

Welcome to the Briskk AI Speech-to-Text Assignment! 🎤
This challenge will test your AI integration, API development, and problem-solving skills through a structured sequence of tasks.

🚀 Your Goal: Build a real-time, noise-resilient voice-based search assistant that: ✅ Converts voice input (audio file or live mic input) into text.
✅ Suggests smart search autocompletions based on user intent.
✅ Handles noisy background audio and improves speech accuracy.
✅ Supports real-time speech-to-search via WebSockets.

---

📋 Assignment Structure

To ensure a smooth progression, complete each task in sequence:

🔹 Task 1: Speech Recognition API (Baseline)

✅ Implement a FastAPI service that:

• Accepts an audio file and converts speech to text using OpenAI Whisper or Mozilla DeepSpeech.
• Returns JSON output { "text": "<transcribed text>" }.
• Test Input: sample_data/clean_audio/sample_english.wav
• Expected Output: "Find me a red dress"

📌 API:

POST /api/voice-to-text
Content-Type: multipart/form-data

---

🔹 Task 2: Handle Noisy Audio (Advanced AI Processing)

✅ Enhance speech recognition by:

• Filtering background noise using RNNoise, DeepFilterNet, or PyDub.
• Comparing accuracy with and without noise removal.
• Test Input: sample_data/noisy_audio/sample_noisy.wav
• Expected Output (after denoising): "Find me a red dress"

📌 Evaluation Criteria:
✔ Speech accuracy before vs after noise removal.
✔ Processing time must remain <1s.

---

🔹 Task 3: Smart Search Autocomplete (AI Ranking)

✅ Implement an API that:

• Suggests relevant results based on user intent & previous searches.
• Ranks results dynamically based on popularity & trends.
• Test Input: "find me"
• Expected Output: [ "find me a red dress", "find me a jacket" ]

📌 API:

GET /api/autocomplete?q=find+me

📌 How to Improve?

• Store previous searches in Redis for ranking.
• Use AI embeddings (OpenAI or BERT) for better matching.

---

🔹 Task 4(optional): Real-Time Speech-to-Search (WebSockets)

✅ Upgrade the system to process live speech queries via WebSockets:

• Accept real-time audio streams.
• Continuously transcribe & autocomplete results dynamically.
• Test: Use a live microphone input.

📌 API WebSocket:

/ws/speech-to-search

✔ Bonus: Deploy the system using Docker & AWS Lambda.

---

🔬 Test Cases (For Self-Validation)

Test Case	Input	Expected Output
Speech Recognition	sample_data/clean_audio/sample_english.wav	"Find me a red dress"
Noisy Speech	sample_data/noisy_audio/sample_noisy.wav	"Find me a red dress"
Autocomplete Query	"find me"	["find me a red dress", "find me a jacket"]
Live Streaming	Microphone	Real-time suggestions

📂 All sample audio files are provided in sample_data/.

---

🏗️ Setup & Running Instructions

1️⃣ Install Dependencies

pip install fastapi uvicorn openai-whisper soundfile numpy scipy

2️⃣ Run the API

uvicorn src.main:app --reload

3️⃣ Test API

• Open Swagger Docs → http://127.0.0.1:8000/docs
• Upload sample_audio_english.wav and check transcription accuracy.

---

🚀 Submission Guidelines

📌 Fork this repo & create a new branch candidate-<yourname>.
📌 Push your implementation & submit a Pull Request (PR).
📌 Explain your approach in a README ( Document trade-offs (e.g., why Whisper vs. DeepSpeech, Redis vs. Pinecone for ranking)). 📌 **Good to have - A deployed version **.

For questions, contact us at: wizard@briskk.one

---

📩 Contact & Discussion

📢 Have questions? Drop an email at wizard@briskk.one 🚀