AI Voiceover and Music: Can Local TTS + Music Generation Replace Online Services? ​

Written in September 2026. AI technology moves fast, so the model capabilities, prices, and hardware requirements mentioned here reflect the situation at that time and are for reference only.

If you make short videos, podcasts, or audiobooks, or need a lot of voiceover, you're no stranger to AI voice and music tools. Voiceover services charge by the character and music services by monthly subscription — at high volume that becomes a running expense.

The good news: audio is probably the area where local AI and online services differ least. For Chinese TTS, local open source models are already very mature; for music generation there are several open source models, each with its own character. And audio demands far less hardware than image or video, so the barrier to entry is low.

Today we go into detail: can local TTS and music generation really deliver, which online services can they replace, and what scenarios suit them.

The conclusion first ​

Speech synthesis (TTS): local can basically replace online services. Chinese output is good, there are many models to choose from, and voice cloning works too. Unless you have extremely demanding requirements for timbre, local TTS is entirely sufficient.

AI music generation: local is close to commercial level. ACE-Step 1.5 is the strongest all-rounder — officially described as sitting between Suno v4.5 and v5. Beyond that there are several specialized models: for background music, for sound effects, for covers, for generating sheet music — each with its own strengths. For most creators, background music, short-video soundtracks, and demo work are all covered.

And the hardware barrier in both areas is low — 4GB of VRAM is enough to run them, which most people's computers can manage.

Local TTS: how far can it go? ​

Which TTS models are available locally? ​

魔当 alone offers a dozen or so TTS apps. The mainstream ones include:

Extra note: besides TTS speech synthesis, there are also local speech recognition (ASR) tools such as FunASR and Whisper-WebUI, which can be used for subtitle generation and speech-to-text. Combined with TTS, they let you build a complete voice workflow.

How does it compare with online services? ​

The most prominent name internationally is ElevenLabs — widely regarded as the top tier of commercial TTS, with very natural timbres and strong emotional expression, plus voice cloning and multilingual support as its signature capabilities. The Creator plan is $22/month (about ¥157), billed by credits, so costs climb fast at high volume. Chinese isn't really its home turf, though — its edge is clearest in English and other languages. For Chinese-language content, local models actually fit better.

There are also quite a few online TTS services in China, such as iFlytek voiceover, Tencent Zhiying, Doubao voiceover, and Tongyi Tingwu. The quality really is good, but most charge by character count or duration.

Where local TTS wins:

• Unlimited generation: no character limits — voice as much as you want. Especially good value for long videos and audiobooks

• Voice cloning: use your own voice or a specific timbre; a few minutes of sample is enough to clone it. This is a premium feature in many paid services

• Privacy: scripts and voices stay local, so sensitive content can't leak

• Free: the models are open source, with no monthly fee

• Batch processing: you can voice and generate in batches, ideal for high volume

Where local TTS loses:

• Installation and tuning: you have to deploy the models and adjust parameters yourself, which takes some effort

• Top-tier timbres: some commercial services' custom voices really are better, but ordinary creators won't need them

• Emotional richness: for very fine-grained emotional expression, online services may be slightly ahead

But for most creators — short-video voiceover, tutorial narration, audiobooks, podcasts — local TTS quality is already completely sufficient.

Voice cloning: can you do it locally? ​

Yes, and it works well.

GPT-SoVITS, CosyVoice, VoxCPM2, and MOSS-TTS 1.5 all support zero-shot or few-shot voice cloning — with only a few seconds to a few minutes of audio sample, you can clone a voice with high similarity. VoxCPM2 also lets you control emotion, speaking rate, and expressive style during cloning.

In online services this is usually a premium feature, charged per use or gated behind a higher membership tier. Locally, you can clone as many voices as you like, completely free.

Good scenarios for it:

• Voicing content in bulk with your own voice

• Giving different video characters different voices

• Making multilingual content with the same timbre across languages

• Character voices for audio fiction

AI music generation: which local models are there? ​

Many people think of Suno when they hear "AI music", but there are already quite a few open source local music generation models, each with its own character. Here's a rundown of the mainstream ones.

Local music models at a glance ​

Let's look at a few of the highlights.

ACE-Step 1.5: the strongest all-rounder ​

ACE-Step 1.5 is an open source music generation model jointly developed by StepFun and ACE Studio, and it sits in the top tier of open source music models today.

Key features:

• Quality close to Suno v4.5 through v5 — that's the official description; actual listening impressions vary by person, but the overall quality really is at the top of the open source field

• Supports 1,000+ instruments and genres — pop, rock, electronic, classical, Chinese traditional, rap... basically all of it

• Runs on 4GB of VRAM — an extremely low barrier; almost any computer with a discrete GPU can run it

• Supports personalized style training — train a style model on your own music material

• Includes stem separation — split a song into separate vocal, accompaniment, and drum tracks

Suited to: creators who need high-quality music generation, want complete songs, and have requirements for arrangement.

HeartMuLa: an edge on Chinese songs and Eastern styles ​

HeartMuLa is another open source music generation model, and it performs well on Chinese songs and Eastern styles (Chinese traditional, Japanese-style).

Features:

• Good Chinese lyric generation — a better grasp of Chinese semantics

• Distinctive Eastern styles — worth a try if you make traditional Chinese-style content

• Lyrics + style generation — input lyrics and a style description and it generates

Suited to: creators making Chinese-language content, traditional-style videos, or anything needing Chinese lyrics.

Stable Audio 3: the background music and sound effects specialist ​

Stable Audio 3 leans toward instrumental music and sound effect generation. It isn't great at songs with vocals, but it's very well suited to background music and sound effects.

Features:

• High-quality background music — instrumental, ambient, and film-score styles all come out well

• Accurate duration control — ask for 30 seconds and you get 30 seconds; ask for 2 minutes and you get 2 minutes

• Strong sound effect generation — it can produce a wide range of ambient and instrumental effects

Suited to: short-video BGM, livestream background audio, film scoring, and creators who need sound effects.

MOSS-SoundEffect-v2.0: sound effects specialist ​

MOSS-SoundEffect-v2.0 focuses on sound effect generation and is a good helper for video editing, game development, and animation production.

Features:

• Wide variety of effects — ambient, mechanical, natural, electronic, and more

• Can also generate background music — not just effects; instrumental music too

• Lightweight and efficient — fast generation with modest hardware demands

Suited to: video editors, game developers, animators — anyone who needs large volumes of sound effects.

SoulX-Singer: the AI cover powerhouse ​

SoulX-Singer doesn't generate music from scratch — it re-sings an existing song in a different voice, that is, singing voice conversion (SVC).

Features:

• High-quality voice conversion — swap singer A's voice for singer B's while keeping melody and emotion

• Trains from few samples — a few minutes of dry vocal is enough to train a timbre

• Supports multiple styles — pop, folk, bel canto, and more are all worth trying

Suited to: music lovers, video creators, and anyone who wants to make covers but isn't a strong singer.

NotaGen: generating sheet music ​

NotaGen doesn't generate audio — it generates scores in ABC notation.

Features:

• Generates editable sheet music — not a final audio file but a score you can import into other software and keep editing

• Good for music study — people learning arrangement or music theory can use it as reference

• Lightweight — it uses almost no resources

Suited to: music students, arrangement hobbyists, and creators who need scores.

How does it compare with online music services? ​

Suno is currently the most popular online AI music generation service; the Pro tier is $10/month (about ¥72) with a daily generation limit. There are similar music generation services in China too.

Where local wins ​

• Free, unlimited generation: no generation caps — make as much as you want. Producing BGM in bulk or trying out styles feels great

• More model choice: you're not locked into one vendor's model; pick per task — Stable Audio 3 for BGM, MOSS for sound effects, SoulX-Singer for covers

• Train your own style: ACE-Step and others support training your own musical style

• Stem separation: generated music can be split into stems for easier post-production

• Privacy: music and material all stay local

Where local loses ​

• Overall level still trails the latest Suno: ACE-Step 1.5 sits between v4.5 and v5, which means there's still a gap to the newest version

• Songs with lyrics: lyric generation quality still trails the top online services

• Installation and tuning: you have to set up the environment, one more step than an online service

• You have to pick the models yourself: having many models is both an advantage and a disadvantage — beginners may not know which to choose

But if all you need is background music, short-video soundtracks, livestream BGM, sound effects, or demo work, local music models are entirely sufficient. And the combination of a different model for a different scenario is actually more flexible than any single online service.

Hardware requirements: audio has the lowest barrier ​

Compared with image and video, audio is far friendlier to your hardware.

Macs with M-series chips can run all of this too. TTS and music generation don't depend on CUDA as heavily, so the Mac experience isn't far off an NVIDIA card.

This is why we suggest that people who want to try local AI start with audio — the lowest barrier, and the easiest place to get positive feedback.

Who is it for? ​

Especially well suited to local ​

• Short-video creators: voiceover, BGM, and sound effects every day; high volume that online quotas can't cover

• Audiobook and podcast creators: long-form audio, where per-character online services get expensive fast

• Content operation teams: multiple accounts and multiple content types need batch production

• Music lovers and independent musicians: making demos, experimenting with arrangement, exploring styles without being limited by subscription fees

• Video editors: you need lots of BGM and sound effects, and local generation is there whenever you want it

• People who need voice cloning or covers: create in bulk with your own voice or a specific timbre

• Educators: voiceover for courseware and teaching audio requires a lot of generation

• Music and broadcasting students: ear training, arrangement, and voiceover practice — unlimited local use with no cost worries

• Language teaching institutions: generate multilingual listening material and pronunciation demos, with data never leaving campus

People who can keep using online services ​

• Occasional users: a few voiceovers and a couple of songs a month — you won't even use up the free quota online

• People chasing top-tier quality: professional music production with extremely high quality demands

• People who don't want to touch technology at all: you just want to upload text and get audio, without installing anything

How to choose? A simple way to decide ​

Do the math and it becomes clear:

If you use an online TTS service, how much do you spend a month?

• A few dozen yuan → doesn't matter; use whatever is most convenient

• ¥100-200 or more → local is better value, with no limits

If you use online music or sound effect services, how much do you spend a month?

• Once in a while → online is fine

• You often need BGM and sound effects → worth installing locally, and you get several models to choose from

More importantly: audio has the lowest hardware barrier. Your GPU may not be up to video generation, but TTS and music run on most computers. The cost of trying it is very low.

Where do you start? ​

If you want to try local audio AI, here's how to get started:

1. Try TTS first — the lowest barrier and the most practical. Start with Qwen3-TTS (lightweight) or IndexTTS 2 (strong emotion control)
2. Try voice cloning — use GPT-SoVITS, CosyVoice, or VoxCPM2; record a few seconds of your own voice and see the result
3. Then try music generation — start with ACE-Step 1.5 and generate a few background tracks in different styles
4. Expand as needed — need sound effects? Install MOSS-SoundEffect. Want covers? Install SoulX-Singer. Making podcasts? Try SoulX-Podcast
5. Build your own audio workflow — script → voiceover → music → sound effects, all completed locally in one pass

With 魔当 you can install every TTS and music app mentioned above in one click, without wrestling with environments and dependencies yourself.

Audio is probably the easiest area of local AI to fall into — good results, a low barrier, obvious savings. If you make content, it's genuinely worth an afternoon to try.