How to Choose the Right AI Voice

Narrow the List Before You Listen

304 voices is not a shortlist, it is a catalog. The voices load from the engine on request rather than sitting hardcoded in the page, so the list reflects what the engine currently serves and can change over time. Auditioning all of them is wasted effort, because most are disqualified by one or two constraints before you hear them.

Apply these filters first, in this order:

1. Language. The interface maps 48 languages to the engine's language_boost setting, which accepts 41 enumerated values including auto. A voice that does not match your script's language is not a candidate, no matter how good it sounds.
2. Register. Decide whether the delivery should read as a person speaking to camera, a narrator outside the story, or an announcer promoting something. That choice eliminates a large share of the catalog.
3. Gender and age impression. Not an absolute filter, but if the brief already specifies a speaker, matching it saves a full listening round.
4. Accent. A hard filter when a neutral accent is not acceptable.

Only after these four does timbre come into play. Emotion, format, and sample rate can adjust delivery quality; they cannot make a mismatched register sound intentional.

Which Voices Fit Which Job

Different production contexts put different demands on a voice. The table below maps the four most common ones to the attributes that matter, and the voice characteristics that tend to satisfy them. These are starting points rather than rankings, because no voice is universally better and the catalog changes as the engine updates.

Narration and documentary

Narration is where an over-performed voice becomes tiring fastest, because the viewer has no visual break from the voice. Listen for a voice that stays level across a full paragraph without flattening into monotony. Test it with the longest, driest sentence in your script, not the opening line. The opening line is where every voice sounds its best.

Advertising

Ads are short, so the ceiling on per-generation length is rarely the constraint. What matters is forward motion. A voice that reads a call to action as if it were handing over an invoice can be lifted with the emotion settings described below, but a voice with inherently soft projection often cannot. Check the read at the volume you will actually mix at, not at full volume.

Character and drama

Cast auditions have an additional constraint: two characters must not sound alike. The fastest way to test this is to generate the same line through two candidates and check whether a listener who was not told which is which could tell them apart. If the voices blur, add emotional separation instead of reaching for a more distinctive timbre, which is the scarcer resource. For serialized production, see the short drama workflow.

Tutorials and explainers

Instructional content is dense with numbers, units, abbreviations, and product names. Audition the technical terms, not the prose. Some voices handle digits and symbols well; others round them off or drop trailing syllables. If your material is heavy on them, generate a stress-test passage and listen specifically for misreads before committing.

How to Test a Voice Properly

A bad audition method produces bad decisions. Use the same script for every candidate, keep the settings identical between runs, and change only one variable at a time.

1. Use one fixed script of 40 to 80 words. Long enough to reveal drift, short enough to repeat.
2. Include your actual opening line. The first five seconds decide whether the rest gets heard.
3. Include one number, one abbreviation, and one proper noun. This is where auditions usually fall apart.
4. Include one sentence with punctuation you rely on. If you use ellipses or em dashes for pacing, test them explicitly.
5. Keep emotion and speed at their defaults during the first pass. You are choosing a voice, not a configuration.

Listen a second time on the device that matters. A voice that sounds thin on studio headphones is often fine on a laptop speaker, and course and video audio is mostly heard on phone speakers.

You can run this whole process in the text-to-speech workspace before committing to anything. Synthesis is capped at 10,000 characters per generation, so keep the audition script well under that.

Choosing and Tuning Emotion

The engine accepts nine emotion values: happy, sad, angry, fearful, disgusted, surprised, calm, fluent, and whisper. These are categorical, not a slider, and they are not all interchangeable with each other or with a pitch or speed control.

Two rules that prevent most emotion problems

Do not stack strong emotion with a strong voice. A highly expressive voice set to angry frequently reads as shouting rather than intent, and the volume eats into your mix headroom. If a candidate already carries a lot of character, pull the emotion down before pushing it up.

Change emotion at sentence boundaries, not mid-clause. Emotional state is interpreted across the utterance. Switching inside a sentence tends to produce a visible seam. If a script genuinely needs a mood change, split the generation into separate requests and join them in your editor, keeping room for a small pause.

Choosing an Audio Format

Seven output formats are available: mp3, pcm, flac, wav, pcmu_raw, pcmu_wav, and opus. The right choice depends entirely on what happens to the file next. Choosing well here can remove an entire conversion step from your pipeline.

If you are unsure, use mp3 for anything you will send to someone and wav for anything you will edit. The two telephony formats are narrow: they are mu-law at 8 kHz, intended for systems that expect that exact encoding, and they will sound poor anywhere else.

One caution: lossy formats are lossy every time. Keep your editing master in wav or flac and export mp3 or opus only as the final delivery. Re-encoding an already compressed file for each revision is a common and avoidable source of audible artifacts.

Choosing a Sample Rate

Selectable rates are 8000, 16000, 22050, 24000, 32000, and 44100 Hz. The engine's ceiling is 32 kHz, which is the default; measured output comes back as 32 kHz mono. Requesting 44100 Hz does not produce a 44.1 kHz file, so treat the two highest options as one target.

The reasoning is straightforward. Higher sample rates preserve more high-frequency content, but synthesized speech does not gain audibly useful detail above 32 kHz, and it costs file size at every later step. Lowering the rate is a real trade-off: 16 kHz is audibly duller than 32 kHz on good equipment.

Mono output suits voice work. Keep it mono unless you have a genuine reason to create stereo, and if you must widen it for a mix, do so after the loudness pass.

When to Use Your Own Cloned Voice

Catalog selection answers "which voice." Cloning answers "whose voice." If the requirement is a specific speaker, a consistent personal brand, or your own narration voice, auditioning catalog entries is the wrong step, because none of them is the voice you actually need.

Reference audio requirements, which catch people out:

• Duration between 10 seconds and 5 minutes
• File size at or under 20 MB
• Format mp3, m4a, or wav

There is also a lifecycle detail that matters for planning: a cloned voice that has not been used in a real generation within 7 days is deleted. If you clone a voice for a project that starts next month, treat the clone as expiring and re-create it when you begin, rather than assuming it will be waiting.

The voice cloning guide covers reference audio selection in more detail, and the cloning page is where you run it.

Keeping a Voice Consistent Across a Project

Consistency is mostly a matter of recording your decisions, not of tuning harder. For any project that will span more than a handful of generations, keep a short spec and reuse it verbatim:

• The voice ID, copied exactly rather than re-selected by name
• The language setting, including whether auto is being used
• The emotion value per section type
• The output format and sample rate
• Any speed or pitch adjustments applied afterward

Two habits cause most drift: re-picking a similar voice partway through, and leaving an emotion setting over from the previous generation.

When you have finished the first segment, listen to it again before continuing. Fixing the settings on segment one is far cheaper than regenerating forty segments at the end because the opening does not match.

Volume is the one variable you should set after generation rather than before. Synthesized clips often differ slightly in perceived loudness between runs, so normalize the finished set as a batch instead of chasing individual files. Common questions about limits and formats are answered on the FAQ page.