Last Wednesday at 11 PM, I stared at my editor for 8 minutes — all for one voiceover line

I had a 2-minute short video. Script done. Visuals cut. All that was left was AI voiceover. I was using one of the mainstream voice tools, and generating a 15-second line took 8 minutes — server was queuing that night.

I thought: "If only I could say a sentence and have AI read it out instantly."

nari-labs got voice generation down to 50ms response

Last week, nari-labs posted on their blog: they got an open-source voice model called Qwen3-TTS responding in under 50ms — that's 0.05 seconds, faster than a blink.

My friend Ajie — an indie audiobook creator in Hangzhou — was paying 200+ yuan a month on ElevenLabs. After switching to running Qwen3-TTS locally, his costs dropped to nearly zero. And because it's fast enough, he started trying a "say one line, generate one line" workflow — recording a whole book dropped from 3 days to 1.

I also got stuck on local deployment — first time I ran the model, the laptop fan went wild, temps hit 90°C. Later I found out you need at least a 16GB VRAM GPU, or just use cloud to start.

What it costs to replicate this today

Money: 0 yuan. The model is open-source and free. Running it locally means no API fees; cloud runs cost a few cents per clip.

Time: If you've never used an AI voice tool before, expect 1-2 hours from download to your first audio output (you'll probably install some things, watch a tutorial).

Tech barrier: It's open-source code, so you need to use Hugging Face — an "AI model marketplace" — to download the model. If you don't want to mess with it yourself, paying a few dozen yuan on Taobao for someone to set it up also works.

First step: Go to huggingface.co, search "Qwen3-TTS", find the nari-labs page, click in and listen to the Demo first. Hear how it sounds before deciding whether to install it.

My advice for three types of people

If you're just starting out (haven't made content yet): It's fine to skip this for now. Use Jianying or iFlytek's built-in voice features — that's enough. Once you're actually voicing 10+ videos per month, then consider switching to TTS.

If you have 1-2 stable clients: If you're on ElevenLabs or similar paid tools, try Qwen3-TTS open-source as a backup — you can save a bundle. But if your client cares about a specific voice, confirm you can replicate their voice first.

If you're scaling up (team of 3+, dozens of audio pieces per month): I'd seriously dig into this sub-second response feature — it lets you build "real-time voice interaction" products, like AI customer service in livestreams or automated phone replies. Still early, but someone's going to grab this first.