[ HÄÄL ]

[ The beginning ]

HÄÄL began with the aim of finding out whether Estonian speech synthesis can be done with a smaller, resource-efficient model that can also be used outside a large cloud infrastructure.

The first experiments focused on adapting an existing TTS architecture to Estonian. The results obtained with a small amount of training material showed that the chosen architecture is able to learn the basic pronunciation and intonation patterns of Estonian speech.

That became the project's first technical starting point. A working model was not a goal in itself, but a basis from which to start assessing quality, data requirements, model size and the possibilities of local use.

[ The Estonian language ]

A separate tokenizer was built for Estonian, and the training material was cleaned and made consistent. The first checked training base consisted of roughly 80 hours of Estonian speech.

The aim of this stage was to verify whether the architecture can learn Estonian pronunciation, rhythm, word forms and intonation stably enough to justify continuing the work.

The results confirmed that the technical direction works, but at the same time showed the need for larger and more varied training material, a consistent evaluation method, and text pre-processing suited to Estonian.

[ More data ]

In the next stage the training material was expanded considerably. Estonian speech corpora available through Kõneveeb were added to the checked material, and the training base grew to more than 600,000 audio clips.

A larger volume of data lets the model learn more speakers, speaking rates, sentence structures and pronunciation variants. At the same time, checking the quality, the origin and the terms of use of the data becomes more important.

In the next stage of development the training material has to be audited systematically, sources with different usage rights separated, and the share of our own checked corpus increased. This is necessary so that future use and distribution of the model does not depend only on data sources that are suitable for experimentation.

[ 250,000 steps ]

In September 2026 the first larger training run reached 250,000 steps.

The results of this checkpoint confirm that the model can synthesise intelligible Estonian speech and use different reference voices. One successful training run does not yet amount to a finished speech synthesis system, however.

The work from here requires comparing checkpoints against a fixed test set, carrying out controlled listening tests, and separating the model's own errors from problems caused by text pre-processing, training data or the quality of the reference audio.

Training will continue from a selected checkpoint with additional, better-checked material. Roughly 150,000–250,000 further steps are planned, but the length of the training will be decided by measured quality rather than by a step count fixed in advance.

[ A voice is not only a model ]

A working neural network is only one part of a speech synthesis system.

For practical use, HÄÄL needs a separate text pre-processing layer that converts written text into a naturally pronounceable form before synthesis.

This covers, among other things, numbers, dates, times, sums of money, percentages, units of measurement, web addresses, email addresses, abbreviations and technical notation.

Developing and testing that layer is a separate language and software engineering task, and it is not solved by training the TTS model for longer.

21.5 °C

twenty-one point five degrees Celsius

https://siseveeb.voco.ee

siseveeb dot voco dot ee

[ Longer text ]

Under laboratory conditions it is easy to test individual sentences. In practical use the system also has to handle paragraphs, articles and longer documents.

For that, a segmentation and joining layer is being developed: it splits the input into logical parts, synthesises them separately and joins the result into continuous audio.

The solution has to preserve the pace, the pauses and the loudness between sentences, and reduce the risk of skipped words and interrupted synthesis.

Handling longer text therefore needs its own development and evaluation, outside the training of the model itself.

[ Voice transfer ]

HÄÄL supports the use of a reference voice, but the quality and the limits of that feature still need systematic evaluation.

Reference audio of different lengths, noise levels, recording equipment and speaker types has to be tested, and it has to be measured how stably the speaker's characteristics are preserved across different texts.

Alongside the technical evaluation, the right to use a voice has to be resolved separately. The licence of the training corpus, the right to distribute the model and the right to use a particular person's voice are not the same question.

The next stage therefore covers both the technical evaluation of reference voices and clear documentation of usage rights and licences.

[ Speed ]

The current development version still uses the larger 24-layer training model.

The first results with INT8 quantisation show that synthesis in real time or faster is achievable on the hardware in question.

These measurements were nevertheless made with the development model, and they are not treated as the performance figures of a final production version.

In the next stage speed has to be measured on different processors and GPUs, quantisation methods compared, and the minimum practical hardware requirements determined.

[ A smaller model ]

One of the main aims of HÄÄL is to reduce the computational cost of speech synthesis.

The current 24-layer model serves as a teacher model and as a technical reference point. The next stage of the research is to distil a smaller student model from it, for example a 6-layer one.

A smaller model is not judged by file size alone. The comparison measures speech quality, synthesis speed, memory requirements and the ability to generalise.

Only on the basis of those results can it be decided which variant of the model suits a server, an ordinary workstation, or a local device with fewer resources.

[ Continued training ]

250,000 steps does not mean that training is over.

For the next stage, a checkpoint from the training so far is selected according to listening tests and technical measurements.

Training continues from the selected model with additional and better-checked training material.

According to the current plan, the next training stage may add roughly 150,000–250,000 further steps.

The exact length depends on whether quality keeps improving, and on which areas still show errors.

The aim of continuing training is not to reach the highest possible step count, but to improve those properties that evaluation still shows to be weak.

[ Data and rights ]

An important part of the next stage of development is putting the training data in order and growing our own corpus.

For material obtained from public and existing corpora, the source, the licence and the permitted use have to be documented.

The right to train a model does not automatically mean the right to distribute the data, to publish the model, or to offer a commercial service based on it.

Our own checked corpus makes it possible to control recording quality, the composition of the texts, pronunciation cases and the distribution of speakers, and gives a clearer basis for later use of the model.

These are treated separately:

• the licences of the training data
• the consent of the speakers
• the usage rights of reference voices
• the terms under which the model itself may be published

[ Evaluation ]

The listening tests so far have mainly guided development. In the next stage, evaluation has to be standardised.

For that, a fixed test set is created which is not used in training, and different checkpoints and model versions are compared on the same inputs.

Besides ordinary text, the test set has to cover names, numbers, dates, abbreviations, loanwords, technical terms, long sentences and other more difficult cases.

In addition to internal evaluation, the aim is to involve independent testers, so that the quality of the model is not judged by the development team's listening alone.

[ A complete system ]

HÄÄL is not designed as a standalone demonstration.

Speech synthesis is one part of a longer technical chain. The follow-on project KUULA adds an Estonian speech-to-text (STT) input layer, and the connecting layer being developed between them makes it possible to link speech recognition to a language model and to speech synthesis.

The longer-term goal is the chain:

KUULA → LLM → HÄÄL

For that, HÄÄL has to be usable through a clear API, with a stable input and output format and documented behaviour.

The next stage of development therefore covers not only the model, but also the API, the integration layer and the runtime environment.

[ Where it stands ]

HÄÄL has reached the point where the core technology has been shown to work for Estonian, but the base technology as a whole is not yet complete.

The current development model is the basis for the next stage, which centres on:

• putting the training data and licences in order
• growing our own corpus
• controlled continued training
• standardised quality evaluation
• text normalisation
• handling long texts
• evaluating reference voices and their rights
• distilling the teacher model into a smaller model
• quantisation and performance measurement on different hardware
• API and local runtime solutions

In parallel, work begins on the KUULA speech-to-text (STT) layer and on the connecting layer between the two.

The aim of the next stage is not simply to train the existing model for longer. The aim is to take a working development model to a documented, measurable and usable base component of Estonian speech technology.