MiniMax Audio vs Fish Audio: Which Fits Better Into a Script
Community Discussion · General

MiniMax Audio vs Fish Audio: Which Fits Better Into a Script

Engineer JiangEngineer JiangNov 112026/11/11 302 views

I compared MiniMax Audio with Fish Audio, which I've been using all along. Ran the same task through both — converting a batch of English technical docs into listening material, about thirty-some segments, called from a script. Conclusion at the end.

Signing up isn't complicated, just go through the email flow and you're in. I'm accessing the overseas version on my end, so you'll have to sort out the network situation yourself — there's a separate entry point for mainland China. Pricing is engineer-friendly, charged per character. The Speech-02 tier is roughly 3.5 RMB per 10k characters, the basic plan is $5/month, and there's a bit of free quota each month. I didn't pay the first time — I ran the whole pipeline through on the free quota first, then decided whether to pay. TTS is a subjective listening thing; get it working before talking money. I'd suggest not skipping that order.

The interface is pretty straightforward — a big input box on the left, a column of voices on the right, click one to preview. There are 300+ preset voices; I got bored after scrolling for ten-some minutes and ended up filtering by language and gender. The Chinese and Cantonese voices really do sound good — the pauses and intonation rises and falls don't sound stitched together.

Two things surprised me. First, you can drag PDFs and TXTs straight in and it extracts the text itself, with a 200k character limit per go. I tried it with a 40-some-page spec sheet — the body text came out basically clean, some header/footer residue left over, just delete it manually. Second, voice design: describe the voice you want in one natural-language sentence and it generates a new voice. I tried "middle-aged male, slower pace, explanatory tone" — the result was usable but unstable; running the same description twice gives different results. On language count, the materials say 30+ in some places and 40 in others; I didn't test them one by one — Chinese, Cantonese, and English are enough for me. For long English sentences, clause breaks occasionally land in weird spots, like cutting right before "which" — that's a bit worse than Fish Audio. Technical terms are mostly pronounced correctly, acronyms are spelled out letter by letter, and I couldn't find a toggle for that.

Both the slowness and the cost come from long texts. 200k characters is the limit, but that doesn't mean it's usable. I stuffed in about 30k characters and the wait was long enough to brew a cup of tea. Engineering-wise this means you can't wait synchronously — you have to switch to submitting a task and polling. The API exists, and I'm familiar with the OpenAI-compatible style. Billing scales linearly with characters, so with lots of long texts the bill is plain as day, no room for discounts. My volume is a few hundred thousand characters a month, so the monthly plan works out cheaper, but I haven't fully figured out how the plan's credits convert to characters. Also, the ten-second cloning thing is real, but it involves authorization and review for real people's voices — the process is more of a hassle than the tech itself.

It suits Chinese- and Cantonese-heavy content, reading long PDFs directly, or needing a batch of different voices for multi-role work. It doesn't suit pure-English technical podcast scenarios where clause breaks are critical, nor individual developers pinching pennies per character. My own approach is to keep both — Chinese goes through MiniMax Audio, English stays with Fish Audio, which is clearly stronger on the Chinese front.

2 replies

?
Ctrl + Enter to reply
Old Deng
Old DengNov 11

Voice design gives different results for the same description across two runs—the stability of this experimental design wasn't controlled. Doing multi-character scripts would be very hard to reproduce.

Early Investor
Reply to Old Deng

The same description comes out different twice—multi-role scripts really can't lock versions; but OP says "running twice gives different results" without mentioning how many runs. I want to ask how many runs were tested—isn't concluding after just two runs a bit premature?