Why text to speech sounds robotic

This is the big one and it accounts for most of it.

Compact voices are small, old, and installed by default on every machine. They are built by stitching together recorded fragments, which is why they land hard on consonants, put the stress in odd places, and have that particular flatness across a sentence. Nothing a tool does will fix that, because the tool is not making the sound.

The tell is that it sounds equally wrong on every tool you try, and that a single word spoken alone sounds fine while a sentence does not.

The fix is a free download from your operating system, covered in the voices guide. It takes about two minutes and it is the only step on this page that most people need.

This one is the tool's fault and it is far more common than it should be.

Speech synthesis works on a chunk of text at a time. Prosody, the rise and fall that makes a sentence sound like a sentence, is computed across that chunk. Hand the engine a fragment and it will read the fragment as though it were a complete utterance, with a full stop's worth of falling pitch at the end.

So a tool that splits badly produces speech that sounds like a list of unrelated phrases. The usual culprits: a sentence containing a link or a bold word gets split at the markup boundary into three pieces; a URL like docs.example.com gets treated as three sentences because of the dots; an abbreviation ends a sentence that had not ended.

The tell is that it sounds fine on plain text and falls apart on a real web page, and that the wrongness lines up with links, bold text and abbreviations.

There is nothing you can do about this from the outside. It is the reason Lumotext joins text across inline markup before splitting, and splits on real sentence boundaries rather than on every full stop.

Synthetic voices degrade unevenly as you speed them up. A neural voice at 1.5× still sounds like a person in a hurry; a compact voice at 1.5× stops sounding like language at all.

Most people set a speed once, on something easy, and then never revisit it. Then they hit a dense paragraph at a speed chosen for a news story and conclude the voice is bad.

Set it while reading something representative of your hardest material, not your easiest. Most people land between 1.1× and 1.4× for familiar material and 0.9× to 1.0× for something they are working through for the first time.

A voice that runs on your own machine has a ceiling, and it is not the same ceiling as a server with a graphics card in it. The very largest models are not going to render in real time on a laptop any time soon. That is worth stating plainly rather than pretending otherwise.

It matters less than it sounds for the thing this is built for. When you are reading along with your eyes on the page, the voice is a metronome: it needs to be steady, immediate and out of the way far more than it needs to be beautiful. A local voice starts speaking with no render step at all, which is what makes clicking a sentence to re-read it feel like nothing rather than like an operation.

And it buys you things a ceiling does not take back: no quota, no account, no connection required, and nothing about what you read leaving your browser.