☕ ParlaBot: potrebbe ripeterlo?

(SpeakBot: Could you repeat that?)

🔗 Demo Link: https://parlabot.io 🔗

ParlaBot is a voice-enabled app that gives you real-time feedback on your Italian pronunciation. Speak into your mic, and ParlaBot will transcribe what you said, compare it to a target phrase, and return constructive feedback — powered by modern open-source AI and traditional DSP filtering techniques.

---

First, a Nod to the Past

My first real speech recognition project was my 2007 Master’s thesis — a vowel recognition frontend built using FFTs, Mel filters, and CMU Sphinx. It’s old-school compared to today’s AI toolkits, but this research (not necessarily mine, but those I studied) laid the foundation for the models that power ParlaBot.
More on that here →

---

Fast-Forward to Now

Nearly two decades later, I’ve been studying Italian seriously for three years and wanted to build something that merges:

• Revisiting of my past studies in speech recognition
• Hands-on exploration of modern STT and AI toolkits
• My passion for learning Italian

Entrare il ParlaBot
(Enter ParlaBot)

---

Project Goals

• Build a practical voice-powered Italian pronunciation coach
• Showcase my ability to design, develop, and deploy AI-based microservices
• Reinforce skills in Python, Go, C/C++, and container-based architecture

---

Architecture Overview

ParlaBot is composed of several Dockerized microservices:

1. Frontend UI in React

   • Displays the target phrase from the PhraseService
   • Records mic input and sends audio to the Orchestrator
   • Displays multiple transcriptions and feedback

2. API Orchestrator in Go/Gin

   • Fetches all target phrases from the Phrase Service
   • Fetches all pipelines from the Audio Preprocessing Service
   • Exposes a /transcribe endpoint
   • Concurrently via goroutines:
     • Forwards the user’s audio to each selected preprocessing pipeline
     • Forwards the filtered audio to the STT Service for transcription and scoring
   • (Planned) Routes results to the Feedback service

3. Audio Preprocessing Service in Python/FastAPI + Torch Transformers

   • Accepts .wav audio
   • Runs audio through specified preprocessing pipelines
   • (Planned) Consume/integrate compoiled C++ shared objects filter chains for audio preprocessing from registry

4. STT Service in Python/FastAPI + HuggingFace Language Model Transcribers

   • Accepts filtered .wav audio
   • Transcribes speech using language model, currently only supports wav2vec2-large-xlsr-53-italian
   • Scores the transcription against the target phrase
   • Returns the model, preprocessing info, and transcript
   • (Planned) Add support for multiple models

5. Phrase Service in Python/FastAPI + MongoDB + coqui (with mozilla and personal speaker training files) + Google TTS API

   • Accepts text phrases and TTS speaker and generates audio using TTS
   • (Planned) Tracks user progress

All services are containerized and connected via docker-compose.

Current System Architecture

... Where its going

---

How to Run

git clone https://github.com/richvigorito/parlabot.git
cd parlabot
docker-compose up --build
open http://localhost:3000

---

Roadmap

see milestones for project milestones/roadmaps/issues/etc

---

License

MIT License

---

Want to read this in Italian?

---

Rich Vigorito | Portland, OR | LinkedIn | GitHub