Dialogue · July 28, 2026
Two synthetic voices can be individually convincing and still fail the moment they have to talk to each other. The difference is almost never the timbre.
Single-utterance text-to-speech has been good enough to fool a casual listener for some time. Put two of those utterances in sequence and the illusion usually collapses. The voices remain clean, the pronunciation stays correct, and the result still sounds like a public address system reading a script rather than two people talking. The failure is not in the voices. It is in everything that happens between them.
It helps to be precise about what a listener is actually tracking. Human conversation carries a dense layer of timing information that nobody consciously attends to but everybody notices when it is wrong: how long a speaker waits before answering, whether the answer starts before the question finishes, how the pitch of a turn's final syllable predicts whether another turn is coming.
Continue reading this note
Earlier notes, most recent first
Prosody · June 12, 2026
Getting the phonemes right is a solved-enough problem. Deciding which syllable carries the weight of a sentence is where synthesis still shows its seams.
Prosody is the collective name for the properties of speech that survive when you remove the words: the melody of pitch, the rhythm of syllable durations, the distribution of loudness, and the placement of pauses. It is the layer that turns a correct pronunciation into a plausible utterance, and it is the layer that most reliably reveals a synthetic voice.
Continue reading this note
Data & Practice · May 4, 2026
Model documentation is written to be skimmed and is most useful when it is not. A short guide to the sections that actually constrain what you can do.
A model card is the document published alongside a released model that describes what it is, how it was built, and where it is expected to fail. Reading one carefully takes about ten minutes and saves a great deal of time later, because most integration surprises are documented somewhere in it.
Continue reading this note
Evaluation & Ethics · March 19, 2026
A few seconds of reference audio is now enough to reproduce a speaking voice. The interesting questions are all about record-keeping.
Reference-based voice synthesis works from a short sample: a speaker reads a few seconds of audio and the system produces new speech in that voice. The technical barrier that once made this a specialist activity is gone, and what remains is a set of record-keeping problems that are unglamorous and entirely tractable.
Continue reading this note
Data & Practice · January 27, 2026
Most of the quality difference between two speech systems was decided during data preparation, long before either was trained.
Corpus preparation is the least discussed and most consequential stage of building a speech system. The choices made here are inherited wholesale: a corpus assembled from clean read speech produces a model that reads, and a corpus with inconsistent loudness produces a model that has learned loudness is random.
Continue reading this note
Evaluation & Ethics · November 15, 2025
Speech quality is a perceptual property, and perceptual properties resist the single number that comparison tables want.
The question people want answered is which system sounds better, and the honest answer is that it depends on which failure the listener is sensitive to. Speech quality decomposes into at least intelligibility, naturalness, speaker similarity and appropriateness of expression, and a system can be strong on one while weak on another.
Continue reading this note
All notes in the archive