Lessons from my first two weeks in a new corner of AI
Two weeks into a new role on the text-to-speech side of voice AI and I keep bumping into assumptions I didn’t know I was carrying.
The switch cost
I came from speech-to-text, where the community conversation is about word error rate, language coverage, and realtime latency. The problems are about getting audio in and text out, fast and clean. Benchmarks are well established. Everyone agrees on what good looks like.
Text-to-speech is a different world. The community talks about voice quality, naturalness, emotion control, pricing per character. The benchmarks exist but they’re more subjective because quality is harder to measure than accuracy. Two people can listen to the same TTS sample and disagree on which sounds better. That barely happens with STT.
I keep reaching for the wrong mental model. I’ll think about a problem in terms of transcription accuracy when the actual question is about prosody or speech mark alignment. It’s like learning to drive on the other side of the road.
What surprised me about the community
TTS developers are more opinionated than STT developers. That makes sense when you think about it. STT is a utility - you want the lowest error rate at the lowest latency. TTS is creative. Developers care about which voice they use, how they control emphasis and pacing, what the output sounds like at different bitrates. They have preferences, not just requirements.
The questions I get are different too. With STT, the questions were about edge cases: accents, background noise, domain-specific vocabulary. With TTS, the questions are about craft: how do I make this sentence sound more natural, how do I control the emotional tone, what sample rate works best for telephony. It’s a more nuanced conversation and I like that.
What surprised me about the market
TTS pricing is more varied than I expected. Some providers charge per character, some per second, some per token. Some have tiers based on quality. Some charge extra for voice cloning or emotion control. Comparing costs across providers requires a spreadsheet and a calculator.
A few years ago, the premium providers charged a significant premium because the quality gap was visible. That gap has narrowed. The top models converged on similar quality, so price has become a bigger factor in buying decisions. That’s good for developers but it means the market is more competitive than the STT side, where a few providers dominate the top tier and everyone else fills niches.
What I’m still figuring out
How to write about a domain where quality is subjective. With STT, I could point to a word error rate and say “this is better.” With TTS, I have to say “this sounds better to most listeners in blind tests” which is a weaker claim. I’m learning to lean on concrete comparisons - pricing tables, latency measurements, language coverage - and let the quality aspect speak for itself through demos and examples.
Also figuring out how deep to go on technical detail. TTS has a lot of parameters: sample rate, bitrate, SSML tags, speech marks, voice cloning settings. Too much detail loses people. Too little leaves them without useful information. Finding the line is taking longer than I expected.
Two weeks in. Still plenty to learn.