• 

Susurro

Inspiration

The word Susurro means whisper in Spanish. The original goal was simple: build a better dictation experience for macOS that is fast, private, and works entirely on-device.

While many excellent open-source transcription projects already exist, most focus on solving a single problem—speech-to-text, subtitle generation, or translation. I wanted to build something broader: a unified speech platform that combines all of these capabilities into one native application.

My long-term vision is for Susurro to become the "AI operating system" for spoken content.

---

What We Built

Today, Susurro provides:

• Local-first speech recognition powered by Whisper
• Global dictation for any macOS application
• Audio and video transcription
• Subtitle generation and multilingual translation
• Searchable transcription history
• Background media downloading and processing

For this hackathon, I began extending Susurro with GPT-5.6 to move beyond transcription into language understanding.

Instead of treating speech as isolated sentences, the AI can understand the surrounding context to improve transcription quality, preserve terminology, and produce more natural translations.

---

Future Vision

Speech contains much more than words—it contains context.

One direction I'm particularly excited about is using a lightweight language model to continuously build context while processing long-form content.

For example, when translating a movie, the model could:

• identify characters and their relationships
• remember names and terminology
• understand the story progression
• keep translations consistent across the entire film

The same idea applies far beyond entertainment.

Imagine meetings where the AI not only generates a transcript, but also understands who said what, tracks action items, summarizes decisions, and serves as an intelligent meeting secretary instead of just a speech recognizer.

Ultimately, I want Susurro to become a platform where specialized AI agents collaborate with speech recognition to make spoken information searchable, understandable, and actionable.

---

Challenges

Building a local-first application required balancing performance, memory usage, and user experience.

Handling different media formats, long-running transcription jobs, background downloads, model management, and native macOS integration introduced many engineering challenges.

Another challenge was deciding what should run locally versus what should leverage larger language models. Rather than replacing local inference, GPT-5.6 complements it by providing semantic understanding and contextual reasoning on top of highly accurate on-device transcription.

---

What I Learned

This project reinforced that speech recognition is only the first step.

The real value comes after transcription—understanding context, preserving meaning, organizing knowledge, and helping users interact with spoken information naturally.

Building Susurro has also shown me how powerful the combination of local AI and frontier language models can be. Local models provide speed and privacy, while GPT-5.6 adds reasoning and contextual intelligence, making the entire workflow far more useful than either approach alone.

This hackathon represents an important milestone toward the larger vision of building a comprehensive AI-powered speech platform.

Built With

• api
• apple
• avfoundation
• codex
• core
• engine
• ffmpeg
• generation
• gpt-5.6
• language
• macos
• ml
• natural
• neural
• openai
• recognition
• silicon
• speech
• subtitle
• swift
• swiftui
• translation
• whisper
• whisperkit
• yt-dlp