Three Bake-offs to Find a Voice Cloning Engine That Runs Entirely on a MacBook

说明与用途

一款 macOS 应用可本地克隆父母声音并朗读作业，配合逐词高亮，使用 Qwen3‑TTS‑0.6B（4‑bit，MLX）、Whisper.cpp 和 Kokoro‑82M。

问题

Build a Mac app that clones a parent’s voice locally on Apple Silicon, converts reading homework into audio with word‑by‑word highlighting, and ensures all data stays on the device with no cloud calls or telemetry.

结果

The final app, Elmren Voice, delivers local voice cloning and word‑highlighted audio on macOS 13+ without cloud dependencies, achieving ~0.88 RTF and 1.71 GB model size with 4‑bit quantization.

工作流

• Record 6–10 s of clean reference audio
• Run Whisper to transcribe and trim to sentence boundaries
• Feed trimmed audio and transcript to Qwen3‑TTS‑0.6B via mlx‑audio
• Generate audio with 4‑bit quantized weights
• Use Whisper.cpp to obtain word timestamps for highlighting
• Package audio and timestamps into a self‑contained HTML player

证据

The hard requirement came from the use case: a child's voice samples and reading materials never leave the laptop. No cloud calls, no subscription, no telemetry.

来源

Qwen3‑TTS‑0.6B-Base is Apache‑2.0 and the community reported RTF 0.86 for cloning.

来源

Same model, same machine: Average RTF PyTorch (MPS) ~9, MLX 0.88.

来源

4‑bit quantized weights: RTF 0.63 vs 1.11 for bf16, size 1.71GB vs 2.4GB.

来源

Real‑world throughput on the M1 Pro: ~1–2.5 minutes per 1,000 characters.

来源