Open-Source macOS Audio Studio for Qwen-Audio TTS Built with Gemini 3.8 Flash

Project Overview

The author (Lao Zhang) previously built a web-based audio generation studio for Alibaba's Qwen-Audio-3.1-TTS-Next model. Users found the web version cumbersome and API scripting too complex. Using Gemini 3.8 Flash as the primary coding assistant, the author rebuilt the tool as a native macOS client with SwiftUI and AppKit, adding extensive features. The entire codebase and a ready-to-install DMG are open source.

Core Architecture

Zero external dependencies : No local Python, Node.js, or ffmpeg required.

Local SQLite database for instant storage and retrieval of projects, prompts, and generated audio.

Native macOS integration : CoreAudio for playback, Finder reveal for generated files, native sliders and shortcuts.

API compatibility : The AI scriptwriter backend accepts any OpenAI-compatible API endpoint.

Feature 1: AI Scriptwriter (AI 编剧)

Converts rough ideas into structured, industrial-grade surround-sound scripts. Preloaded scenarios include:

Dual-host tech podcast

15-second product ad

Suspense radio drama

Fantasy tavern voiceover

Given a prompt like "write a 15-second coffee ad with door chime and steam wand sounds", the scriptwriter outputs a script with four sections: [任务] (Task) [声场与环境] (Soundscape & Environment) [音效时间轴] (SFX Timeline) [人物与台词] (Characters & Dialogue)

One click inserts the parsed script into the main creation studio.

Feature 2: Native Creation Studio (原生创作台)

Full macOS-native editing environment:

Quick-tag toolbar : Insert [角色], [对白], [时间戳], [音效], [音乐], [音色] tags; adjustable editor font size.

Acoustic parameters : Output format (WAV, MP3, PCM), sample rate (16000, 24000, 48000 Hz), speech rate and volume via native sliders.

Reference voice support : Upload local audio, trim to ≤30 s waveform segment, use ad-hoc or save permanently to local voice library.

Before generation, a semi-transparent preflight window shows the complete compiled prompt, output directory, reference voice file, and estimated billing count. The request fires only after explicit user confirmation.

Feature 3: Instant Playback & Delivery (即刻试听与交付)

Four-stage visual progress monitor:

参数编译准备 → 提交百炼任务 → 异步声学渲染 → 结果下载与入库

On completion, a delivery card appears with two macOS-specific optimizations:

Direct playback : Native play/pause button streams via CoreAudio, handles Bailian OSS streaming WAV dynamic headers, displays exact physical duration (e.g., 15.0 s).

Finder reveal : One click highlights the actual generated file in Finder for immediate drag-and-drop into Final Cut, WeChat, etc.

Feature 4: Project Library & A/B Blind Testing (作品库与同位置 A/B 盲测)

Every generation — audio, prompt, parameter set — is silently persisted in the local SQLite database. The library supports keyword search, tag filtering, and waveform thumbnails. Inside a project, same-position A/B blind testing lets the user switch between version A and B at the exact same playback timestamp to compare speech naturalness, background music blend, etc.

Feature 5: 42 Ready-to-Use Templates (42 套开箱即用的灵感模板)

Seven scenario categories (Podcast, Ad, Audiobook, Radio Drama, Game Dubbing, Narration, Custom), each with 6 battle-tested templates (42 total). Templates pre-configure character voice settings, soundscape, and SFX trigger points; users only replace theme variables.

Open Source & Future Plans

The project is fully open source. The author invites stars, issues, and pull requests, and plans to add support for more TTS models beyond Qwen-Audio.