LocalText2Voice: Issues and Improvement Suggestions

[Hela-Girl]

1. OmniVoice installation prompt appears on every launch

Every time LocalText2Voice starts, it prompts the user to install the OmniVoice TTS model, even though F5-TTS Russian is already installed and selected.

The application may be failing to detect the currently installed and active TTS model, or it may incorrectly treat OmniVoice as a required dependency.

2. Custom voices cannot be edited

After creating a custom voice, there is no way to edit its settings, replace the audio sample, correct the transcription, or update other parameters.

To make any changes, the user has to delete the voice or create a new one.

It would be helpful to add an option for editing existing custom voices.

3. Voice sample playback cannot be paused or stopped

After starting a voice sample, there are no controls for:

• pausing playback;
• stopping playback;
• restarting playback from the beginning.

The user has to wait until the sample finishes playing.

4. Model downloads fail when the application is installed outside the default location

If the application is installed outside the default folder or on a drive other than C:, model downloads stop working.

The application reports that it cannot find the required files or directories.

The application may be using hardcoded paths instead of the actual installation path.

5. Missing normalization for the Russian letter “Ё”

The Russian text-processing system lacks automatic normalization for the letter “Ё”.

In many texts, “Ё” is written as “Е”, but this can change the pronunciation, stress, and intonation of words.

It would be useful to add context-aware normalization that automatically restores “Ё” where appropriate.

6. One of the voices is sometimes ignored when using voice tags

When male and female voice tags are inserted into the text, one of the selected voices is sometimes ignored during generation.

The interface does not clearly indicate:

• which voices are currently selected;
• the exact tag assigned to each voice;
• the correct format for writing voice tags;
• whether a tag has been recognized successfully.

Suggested improvements:

• Display the exact tags for selected and uploaded voices.
• Add a button for automatically inserting a voice tag.
• Show an example of the correct tag format.
• Warn the user when a voice or tag cannot be found.
• Validate all voice tags before starting generation.

[estebanstifli]

• Fixed the repeated OmniVoice installation prompt and model-path handling outside the default install location.
• Added editing and permanent deletion for imported custom reference voices.
• Added pause, stop, and restart controls for voice previews, plus per-segment playback controls in Generation Review.
• Improved voice markup selection with local voice pickers and direct tag insertion.

For Russian normalization, we tested multiple approaches. None can guarantee 100% accuracy for every Russian homograph or omitted “Ё” case without human context. We kept Silero Stress because it produced the best practical results; it is now an optional, on-demand Russian normalization layer.

[Hela-Girl]

Thank you for the fixes. I really appreciate how quickly you respond to feedback.
I would like to point out two issues that unfortunately still have not been resolved.

1. I still cannot edit or delete previously created custom voices. When I try to edit a voice, the application reports that a required file cannot be found and then suggests installing voice components that are already installed on my system.

2. Every time the application starts, it still prompts me to install OmniVoice, even though I already have another TTS engine installed and in use. It would be great if the application could detect an existing TTS engine and avoid suggesting OmniVoice when the user is already using another one.

[19:26:58.339] Audio Mix: waveform data ready for drawing. [19:26:59.905] Обнаружено голосов Piper: 3. [19:27:07.604] Установка OmniVoice отложена. [19:27:30.989] Gallery voice file not found: D:\LLM\LocalText2Voice\data\data\voice-gallery\imported\f5_russian\female_chistyakova_2.wav

[Hela-Girl]

3. There is also an issue with switching between TTS models. If I simply select another TTS model in the interface but do not install or download it, the Voices section immediately starts displaying voices associated with that model, even though a different model is actually installed and available on my system.

In my opinion, voices belonging to a model should not become available until that model has actually been installed. Alternatively, the interface should clearly indicate that the selected model is not installed yet.

[Hela-Girl]

4. There is also an issue with unloading the model from memory. After closing the application, the related process sherpa-onnx-ws-win32-x64.exe continues running and uses around 750 MB of RAM.

As I understand it, this is a separate process or runtime used by the application for model-related operations. After LocalText2Voice is fully closed, I believe this process should also terminate automatically.

Sorry to bother you again, Esteban, but I thought it was worth reporting this as well.

[estebanstifli]

@Hela-Girl
Thank you for reporting this.

These issues have now been addressed in the latest version:

Fixed legacy custom voices whose paths could incorrectly become data\data, preventing edit/delete.

Fixed the repeated optional OmniVoice prompt when another TTS engine is selected.

Voices from engines that are not installed are now clearly unavailable until their engine is installed.

Regarding sherpa-onnx-ws-win32-x64.exe: it is not bundled or started by LocalText2Voice. Please share its full file path/command line from Task Manager, as it likely belongs to another installed application.

Please update to the new version and let me know whether it resolves everything on your system.

[Hela-Girl]

Esteban, I have found a few more issues.

I currently have two TTS engines installed, but there are no available voices shown for text generation.

I also noticed another issue: simply clicking on a TTS model in the list immediately makes it appear as the selected engine in the bottom-left corner, even if that model has not actually been installed. I think only an installed and available engine should become active.

Additionally, I still cannot open or edit the custom voices I created.

If possible, could you please take a look at these issues and fix them? I would really appreciate it.

[estebanstifli]

Thank you — I identified the root cause.

Older versions could save the managed data folder itself as the storage location, which resulted in a legacy data\data layout. The new path validation prevents this for new installations, but existing installations also need their voice gallery database, imported reference audio, and model paths migrated to the canonical location.

I am implementing a safe startup migration with a progress dialog. It will copy and verify the old assets, update the stored paths, and keep the original folder as a backup.

I am also fixing the engine selector so a local TTS engine cannot become active until it has actually been installed.

[Hela-Girl]

Also noticed that in the latest build, the program freezes on a small text segment. That is, the program responds visually, as if the model has been loaded into memory, but the generation process does not occur. The program simply does something, but no result is visible.

Maybe the problem is with: Silero Stress
sys.stdin.read()

log:
[18:20:29.441] Начало генерации...
[18:20:29.480] Connecting to the shared LocalText2Voice engine host...
[18:20:29.480] Engine host job: 9df00e5d-cc8c-40b1-8889-c8f327f2d2eb
[18:20:29.459] Engine host reusing f5_russian engine from memory.
[18:20:29.466] Using TTS engine: F5-TTS Russian (non-commercial)
[18:20:29.469] F5-TTS Russian model: f5tts_v1_base_v2
[18:20:29.471] Silero Stress: enabled
[18:20:29.474] Validation completed in 8 ms.

I encountered an issue where generation remains stuck at 0% when using F5-TTS Russian with Silero Stress. The Silero processes start but remain idle, GPU usage stays near zero, and audio generation never begins.
Please consider the following changes:
Close stdin after sending the JSON, or use communicate(input=...).
Use readline() instead of sys.stdin.read() to receive a single line.
Prevent Silero from running simultaneously in both the UI and EngineHost.
Add a 30-60 second timeout. If Silero does not respond, continue generation without automatic stress placement.
Write processing stages and errors to the log file. Currently, local_text_2_voice.log remains empty.
Display a separate Applying stress marks stage so generation does not appear frozen at 0%.
I also tested Silero Stress separately. The model loads and processes a test sentence within a few seconds. Therefore, the problem is most likely in the communication between LocalText2Voice and the Python subprocess.

[Hela-Girl]

I found another issue related to the Faster Whisper dependency.

If Faster Whisper is not installed, the application does not allow normal audio generation and requires Whisper to be installed, even when I do not need any text verification or transcription functionality.

In my opinion, Whisper should be an optional dependency. If the user is not using text verification, transcription, or any other Whisper-related feature, speech generation should still work without installing it.

At the moment, users are forced to install

[Hela-Girl]

And there is one more issue, probably the most important one, related to using multiple voices in the same book.

When the text switches from one narrator or character voice to another, the end of the previous voice segment gets cut off. In other words, the previous audio segment ends too early before switching to the next voice, causing part of the last word or the end of the sentence to be lost.

This is especially noticeable in dialogues where the voices frequently switch between characters.

It would be very important to ensure that each audio segment is allowed to finish completely before switching to the next voice, and that the end of the previous segment is not trimmed when the generated audio segments are combined.

Book:

{{lang ru}}
{{voice "Russian Woman"}}
Академия
Бентли Литтл {{play "relax2.mp3" track=music volume=-15db}}
Пролог

{{voice "Russian Man"}}
– Чем займёмся?

{{voice "Russian Woman"}}
– спросил Ван.
{{voice "Russian Man"}}
– Без понятия,

{{voice "Russian Woman"}}
– пожал плечами Курт.

{{voice "Russian Man"}}
– Есть мысли? – Нет.
{{voice "Russian Woman"}}
Двое подростков прислонились к закрытым гаражным дверям. Впереди лежала улица и долгий нудный день.
{{voice "Russian Man"}}
– Может, в баскетбол? – Давай,

[estebanstifli]

Thank you for the detailed reports and screenshots.

F5-TTS Russian freezing with Silero Stress:

I’m going to investigate the Silero Stress subprocess and its communication with the shared Engine Host. I will also review timeout handling, progress reporting, and the fallback behavior if stress normalization becomes unresponsive.

Faster Whisper requirement:

Faster Whisper is optional for normal TTS generation and automatic review. However, it is also required when using the {{play}} markup command.
LocalText2Voice uses Whisper’s word-level timestamps to determine the exact position of words in the generated audio. This allows music and SFX commands to be synchronized accurately with the narration. Therefore, when a project contains {{play}}, Faster Whisper must be installed even if Generation Review is disabled.
I agree that the current error message should explain this more clearly.

Cut endings when switching voices:

Could you please confirm whether this happens with both of your installed TTS engines, or only with F5-TTS Russian?
F5-TTS Russian may be less reliable with very short fragments. It also has a Remove long silences option enabled by default. Please try disabling this option in Settings → TTS Engines → F5-TTS Russian, generate the same dialogue again, and let me know whether the endings improve.
This will help determine whether the audio is being shortened during F5 generation or at a later stage.

[Hela-Girl]

omnivoice - In this ending model, it does not eat, but the text generation speed I demonstrated above took 28 minutes.
F5-TTS Russian - In this model, word endings are truncated only when switching from one speaker to another. However, this model offers significantly faster generation quality and speed. The same segment was generated in approximately one minute.

In settings, the option to delete long pauses has no effect on chewing the ends.

[Hela-Girl]

There is another issue related to F5-TTS and the available voices.

The settings indicate that the application uses the Misha24-10 model based on F5-TTS, and it also states that voice cloning and generated audio cannot be used for commercial purposes.

However, after installing this model, it does not appear in the list of available voices at all. The Voices section still only shows the default Russian Man and Russian Woman voices that come with the application.

So there is currently a mismatch between the settings and the voices actually available in the interface: the application reports that Misha24-10 is installed and used, but its voice cannot be selected anywhere.

[estebanstifli]

Thanks for reporting this. Misha24-10/F5-TTS_RUSSIAN is the underlying Russian TTS model, not an additional selectable voice.

The “Russian Man” and “Russian Woman” entries are preset reference voices, and both use the installed Misha24-10 model for generation. F5-TTS can also clone a custom voice from a reference audio, but installing the model itself does not add a voice named “Misha24-10”.

I agree that the current wording in the settings can be confusing because it does not clearly distinguish between the TTS model and the available reference voices. I will improve the labels and description to make this clearer.

The non-commercial restriction applies to the Misha24-10 model weights and therefore also to audio generated with the preset or cloned voices through this engine.

[estebanstifli]

Thanks, this is helpful. F5-TTS has known problems with very short segments, including omitted or truncated words (#460, #1085). The speaker switch may be creating a fragment that is too short. Since disabling long-pause removal makes no difference, the ending is probably already truncated during generation.

However, 28 minutes with OmniVoice is very unusual. Could you tell me your GPU model, VRAM, RAM, CPU, and whether CUDA was being used? Please also paste the exact text and speaker markup for that segment so I can try to reproduce both issues on my machine.