How to Choose an AI Voice Generator: 9 Questions

Disclosure: written by the VoiceSmiths team. We sell an AI voice studio, so we are not neutral; this guide is the checklist we would use ourselves, and most of it applies to any tool. Where we quote our own numbers we say so; for other tools, check their own pricing pages.

Choose an AI voice generator by the job you need done, then test it on your own script for fifteen minutes before paying. Most disappointment comes from buying a tool built for a different job, or from judging a voice by its polished demo instead of by your brand names, your dialogue and your longest paragraph.

The short answer

1. What exactly will you make?

Write down the three things you will produce most often in the next month, with their length. "A 30-second product video twice a week" and "a 90,000-word novel" need different tools, and a tool that is excellent at one is often mediocre at the other. If more than one character speaks, you need per-line voice assignment; if it is longer than a few thousand words, you need a long-form mode.

2. Which languages, and do you need a native voice in each?

Many models now speak dozens of languages (Eleven v4 covers 90+), but a voice keeps the accent of the language it was designed or recorded in. A Mandarin voice reading English sounds like a Mandarin speaker reading English. That can be what you want, a character who keeps their voice in every language, or exactly what you don't want, an ad for a Spanish market. If the accent matters, check that the tool has a native voice for each language you publish in.

3. How will you test quality?

Ignore the homepage demo; it was chosen to impress. Test three things of your own:

1. A line full of your words: brand names, product names, acronyms, people and places.
2. A line of emotional dialogue: an argument, a joke, a whisper.
3. Two minutes of narration: listen to the end, because drift in pace, volume or tone shows up late.

Listen on phone speakers and on headphones. Many tools offer a few free generations without an account; on VoiceSmiths it is two short clips a day from the homepage or the text-to-speech page, which is enough for tests 1 and 2.

4. Can you direct the performance?

A good take usually needs more than one try. Check what you can change without rewriting the text: emotion or audio tags (such as [whispers] or [laughs]), speed, stability, and per-line control in dialogue. Newer models follow inline tags; older ones read them aloud. Our emotion tags cheat sheet shows the tags that work.

5. Can you fix a word it keeps getting wrong?

Every TTS model mispronounces something: your brand, a surname, an acronym, a Chinese polyphone. Ask whether you can pin a pronunciation once and have it apply everywhere (a pronunciation dictionary or alias rules) rather than respelling the word in every script. Test it with your five hardest words. See how pronunciation dictionaries work.

6. What does a finished minute cost?

Plans are priced in credits, characters or minutes, which makes them hard to compare. Convert everything to the cost of one finished minute:

• English speech runs at about 850 characters a minute; Mandarin at about 250 characters a minute (our own renders: 1,000 English characters ≈ 70 seconds, 1,000 Chinese characters ≈ 4 minutes).
• Multiply by the per-character price, then add your retakes. A realistic budget is 1.5 to 2 times the final length.
• Check whether monthly credits expire at the end of the month, whether failed generations are refunded, and whether there is a one-time option so you can buy a test without a subscription.

Our numbers, for comparison: speech is 7 credits per 100 characters (4 on the faster Flash model), which comes to about $0.47–0.70 per finished English minute on the monthly plans; failed generations are refunded automatically; and a $3.99 one-time pack holds 150 credits that do not expire.

7. Can you use the audio commercially?

Read the license before you publish. Things to check: whether the free tier allows commercial use (often it does not), whether a voice imitates a real person (avoid it), what the rules are for cloning a voice (yours, or one with written consent only), and whether the platform you publish on asks you to disclose AI audio. Chinese platforms, for example, require AI-generated content to be labelled.

8. What do you get out?

At minimum, a downloadable MP3 or WAV. For video work, timestamps or SRT subtitles save an hour of alignment per video. For production, check for an API, batch generation, and whether you can download stems or separate lines. If you edit in CapCut or Premiere, an SRT that lines up with the audio is worth more than a slightly better voice.

9. Can you pay, and get help, the way you work?

This rules out more tools than quality does. Most Western platforms take cards only; most Chinese platforms take Alipay or WeChat Pay and often require a Chinese phone number. Check the invoice you need, the support language, and whether a real person answers. VoiceSmiths takes Alipay in CNY and cards in USD, with a Chinese and English interface.

A 15-minute test plan

1. Write the three test lines from question 3, using your own script.
2. Generate each in two or three voices on the free tier.
3. Add your five hardest words to a pronunciation rule, if the tool has one, and regenerate.
4. Export once and drop it into your editor with the subtitles.
5. Work out the cost of one finished minute (question 6).
6. Read the license paragraph on commercial use (question 7).

If a tool fails step 1 or step 4 for your content, no feature list will fix it.

Red flags

• "Unlimited" generation for a few dollars a month, with no fair-use terms.
• Voices named after, or obviously imitating, celebrities.
• Cloning anyone's voice from a short clip with no consent step.
• No word on commercial use, or on what happens to unused credits.
• Prices only visible after you hand over a card.

FAQ

What is the best AI voice generator? There is no single best one; there is a best one for a job. For short English voiceovers, quality differences between the leading models are small, so price, editing controls and export matter more. For multi-character work, per-line casting matters most; for Chinese content, native Mandarin voices and Alipay matter most.

Is a free AI voice generator good enough? For testing and personal projects, often yes. For commercial work, check the license: many free tiers forbid commercial use or watermark the audio.

How much should AI voiceover cost per minute? It varies by an order of magnitude between consumer plans and APIs. Convert every plan to the cost of one finished minute, including retakes, before comparing. On VoiceSmiths it is about $0.47–0.70 per English minute on the monthly plans.

Should I clone my own voice? Only if your voice is part of your brand and you will produce a lot. A cloned voice is only as good as the recording you give it; a well-chosen library voice is usually better than a clone made from noisy audio. See our voice cloning tutorial.

How do I compare two tools fairly? Same script, same language, same length, both on their best model, listened to blind. Our head-to-head pages (vs ElevenLabs, vs Murf, vs Speechify) follow that rule, and list the rows where the other tool is better.

Test your own script free, no sign-up →