🌉 Bridge-TRL

Bridge-TRL is a comprehensive, all-in-one Dockerized backend providing streaming capabilities for Translation, Speech-to-Text (STT), Text-to-Speech (TTS), Image-to-Text (ITT), and Text Inflection.

It was specifically designed for GoTRL. While GoTRL acts as the router handling data distribution to workers, Bridge-TRL serves as the ultimate example of how to build high-performance, WebSocket-based endpoints that GoTRL can interact with.

✨ Key Features

• GoTRL Ready: Built natively to seamlessly integrate with GoTRL.
• All-in-One Toolkit: Consolidates 5 powerful tools (STT, TTS, ITT, Translate, Inflector) into a single unified service.
• Lock-Free Streaming: Used a custom atomic ring buffer for the hotpath, ensuring zero-allocation memory reads/writes and ultra-low latency during real-time WebSocket communication.
• EasyTranslate Inside: The translation module is fully powered by my lightweight, memory-optimized EasyTranslate library, handling context-aware text transformations locally.
• Flexible AI Backends: Supports local offline models (Vosk, RHVoice, Ollama) as well as external HTTP APIs or bash scripts to process data.

🐳 Deployment (Docker)

Bridge-TRL provides two deployment strategies depending on your network constraints and available offline models.

Option 1: Local / Offline Build (Tested & Recommended)

Use this option if you face network restrictions (e.g., alphacep servers being blocked) or prefer using pre-downloaded models. It utilizes local instances of rhvoice and vosk.

Make sure your models are placed in the ./assets/ directory before building.

docker build -t bridge-trl -f Dockerfile .
docker run --network host bridge-trl

Option 2: Full Standalone Build

This Dockerfile attempts to download all necessary dependencies and models (including Vosk) automatically during the build process.

⚠️ Warning: I have not fully tested this build. If your network blocks alphacep servers, Docker will fail to download the models and timeout. Use Option 1 if you encounter this issue.

docker build -t bridge-trl -f Dockerfile.full .
docker run --network host bridge-trl

🏗 Architecture & Custom Workers

Bridge-TRL isolates every processing unit into a modular interface. If you want to write your own worker, you just need to implement this simple interface:

package workers

import "net/http"

type Worker interface {
    GetName() string
    Register(m *http.ServeMux)
}

Data Flow

graph LR
    A[GoTRL Router] -- WebSockets --> B(http.ServeMux)
    B --> C{Worker Module}
    C -- Writes --> D(RingBuffer IN)
    D -- Spin-reads --> E[Processing Engine]
    E -- Yields --> A

Loading

📖 Endpoints Reference

All endpoints operate over WebSockets (ws://) to support continuous data streaming.

• /stt (Speech-to-Text): Powered by Vosk. Streams PCM audio bytes and returns recognized text. Handles partial string trimming automatically.
• /translate (Translator): Powered by EasyTranslate. Accepts a setup message (e.g., en ru) or uses auto-detection, then translates text streams on the fly using local key:value dictionaries.
• /tts (Text-to-Speech): Powered by RHVoice. Converts text chunks into audio bytes. Supports fallback to external APIs or local bash scripts depending on your configuration.
• /itt (Image-to-Text): Powered by gosseract (Tesseract OCR). Accepts image bytes over WS and returns the extracted text string.
• /inflector (Inflector): Intelligent grammar corrector. Useful for adjusting grammatical agreement (word forms, gender, cases) of translated text against the original context using AI models (like Ollama).

📜 Licenses

This project includes and links against several open-source libraries. To ensure full legal compliance when copying, modifying, or distributing Bridge-TRL, here is a breakdown of what license applies to each module, what it is used for, and where to find its original terms.

Dependency / Module	License	Used For	Reference Link
Vosk API	Apache-2.0	Core backend for the Speech-to-Text (/stt) worker. Handles PCM audio stream recognition.	alphacep/vosk-api
EasyTranslate	MIT	Lightweight, memory-optimized text translation backend used by the /translate worker.	Votline/EasyTranslate
Go-Audio	MIT	Record, play, and handles low-level manipulation of audio byte streams.	Votline/Go-audio
Gorilla WebSocket	BSD-2-Clause	Provides full-duplex streaming capabilities across all API endpoints (/stt, /tts, etc.).	gorilla/websocket
Gosseract	MIT	Go wrapper around the Tesseract OCR engine, running inside the Image-to-Text (/itt) worker.	otiai10/gosseract
Zap	MIT	High-performance, structured, and leveled logging package applied globally in the router.	uber-go/zap

• License: This project is licensed under MIT
• Third-party Licenses: Third-party licenses/.