Kazakh TTS Data: The Four Decisions That Decide Whether a Corpus Is Usable

Kazakh is spoken by more than thirteen million people, most of them in Kazakhstan, and the speech data recorded in Kazakh for synthesis is comparatively thin. Anyone commissioning a Kazakh speech dataset for text-to-speech work runs into the same four decisions, and each one is cheaper to settle before recording than after.

Most of the Kazakh speech corpora that get discussed publicly were built for recognition, and the two jobs pull in opposite directions: recognition tolerates variation because the decoder averages over it, while synthesis amplifies variation into artefacts a listener hears immediately. A Kazakh speech dataset is usable for text-to-speech when the text layer and the audio layer describe the same language consistently, and that condition is decided in specification, not in post-processing.

The gap is not closed by recording more hours of the same thing. In Kazakh speech projects, four decisions get made before the first hour is recorded, and each one is expensive to reverse later. The Kazakh language page covers the script and dialect picture; the dataset briefs show how the same choices play out in other languages.

Decision one: which script the text layer uses

In Kazakhstan, Kazakh is in the middle of a script transition. Cyrillic remains in wide use and a Latin-based alphabet is being introduced in stages, with full adoption targeted within the current decade. Both scripts circulate at once, and a speech corpus that mixes them is harder to use than either script on its own. The script decision is the first thing that separates a usable Kazakh speech corpus from an unusable one.

The choice does not change the audio. It changes the text paired with the audio, and through that it changes what the corpus can train. A synthesis model learns the mapping between written units and sound. If the written side switches systems partway through a corpus, the model learns two different mappings for the same sounds, and the loss appears as inconsistent pronunciation rather than as an error anyone can point at.

• Decide the script before recruitment, not after recording.
• State it in the delivery contract and in the file naming.
• If both scripts are needed, commission two parallel text layers from the same audio rather than one mixed layer.

Decision two: whether vowel harmony survives transcription

Kazakh has nine vowels and a set of letters that do not exist in Russian Cyrillic: ә, қ, ғ, ң, ү, ұ, і, ө, һ. Vowel harmony governs which affixes a word can take, so replacing one vowel letter with another does not merely misspell a word. It produces a form a native speaker recognises as wrong.

The failure is common in transcription, because several of these letters sit close to ordinary Cyrillic letters on a keyboard and in low-quality fonts. Models trained on the resulting text confuse қ with к, ң with н, and collapse ү and ұ into a single u. The errors are systematic, and every model trained on the corpus inherits them.

Transcription for synthesis has to preserve the original letters. Normalising to a simplified alphabet to make annotation faster moves the cost downstream, where it is much harder to find. The Kazakh TTS brief takes the same point from the language side.

Decision three: where a Russian loanword stops being Kazakh

Kazakh and Russian share the Cyrillic script, and Russian loanwords are embedded in ordinary Kazakh speech to a degree that surprises anyone who has not worked with the language. Off-the-shelf language identification tends to label these loanwords as code-switching, because the character statistics look Russian.

Whether that labelling is correct is an annotation decision, not a modelling one. Two conventions are defensible: count integrated loanwords as Kazakh and reserve the mixed label for clause-level switches, or label every Russian-origin token separately. They produce different training signals and different downstream behaviour.

The convention has to be written down before annotation begins and applied consistently, because annotators working without a written rule disagree with each other, and with themselves a week later.

Decision four: speaker design, which matters more than hours

Buyers usually open with a number of hours. Hours are the easiest quantity to produce and the least informative one to compare. Two corpora of equal length can support completely different models depending on how many speakers they contain.

A small number of speakers recorded for a long time supports voice cloning and high-fidelity single-voice synthesis. A larger number of speakers with shorter recordings supports speaker generalisation and multi-speaker models. High-quality single-speaker Kazakh material is scarce relative to multi-speaker crowdsourced material, so the speaker design should be stated explicitly rather than assumed.

Region belongs in the speaker metadata. Kazakh speakers recruited in the south of Kazakhstan and in the north do not sound the same, and a corpus that mixes them without recording where each speaker came from cannot later be split by region. Recording the city or oblast at recruitment costs nothing; recovering it after delivery is usually impossible.

• How many distinct speakers, as a minimum rather than a target.
• Gender and age balance, if the downstream model needs it.
• Whether the same speaker appears across the whole corpus or only in part of it.

What the recording itself has to satisfy

Decisions about text do not fix problems in the audio. Four properties carry most of the quality:

• Sample rate. Most synthesis pipelines treat 16 kHz as a floor, and higher rates are only worth the storage if the rest of the chain preserves them.
• Noise floor. A recording made close to the noise floor of the room cannot be recovered by processing; the processing removes the noise and part of the speech with it.
• Clipping. Peaks that hit the ceiling are unrecoverable, and are easy to introduce when levels are set by ear.
• Reverberation. A room that sounds pleasant to sit in can still make speech unusable for synthesis, because the model learns the room as part of the voice.

Over-processing is its own failure. Aggressive denoising and compression remove the variation that makes synthetic speech sound human, and the loss is invisible in a waveform inspection. The technical reference explains how the common quality metrics behave when they are applied to processed speech.

A checklist to settle before commissioning

• Script: Cyrillic, Latin, or two parallel text layers.
• Transcription convention for the nine Kazakh-specific letters.
• Loanword rule, written down and applied consistently.
• Speaker count and speaker design, with a minimum rather than a target.
• Sample rate, noise floor and clipping policy.
• Sentence length and segmentation convention.
• Whether the text layer includes punctuation, and how numbers are handled.
• Acceptance test: who listens to how many samples, and against what criteria.

A Kazakh text-to-speech dataset usually fails at the text layer long before it fails at the microphone, and every failure of that kind is traceable to a decision nobody wrote down.

None of these decisions is expensive to make at the start, and all of them are expensive to reverse. A corpus that fails is usually the one where a decision was left to whoever happened to be doing the work that day. For the same reasoning applied to a different class of speech, see conversational speech.