Building a Text-to-Speech Solution for Less Common Languages — Stuck on TTS

Hey everyone,
I’m working on a project aimed at generating text and converting it into speech for languages that are not commonly supported by most TTS tools (basically, anything outside English, Spanish, French, etc.).
The concept is simple: I generate text content based on a given theme or input (this part works fine with GPT models). The challenge is converting that generated text into audio — in a specific less-common language.
Here’s what I’ve tried so far:

• Coqui TTS:Tried multiple versions (0.22.0, 0.9.0, etc.). Ran into constant issues like:
  • No module named 'TTS.tts.train'
  • No module named 'TTS.trainer'
  • Missing or deprecated training modules
• gTTS (Google Text-to-Speech): Limited language support and the audio output quality isn’t good enough for this use case
• Voice cloning attempts: Faced compatibility issues, lack of documentation, or outdated repos when trying to build or fine-tune models on my datase

I’m open to any suggestions, libraries, or approaches that could help me:
- Train or fine-tune a TTS model on my own dataset (for a less-supported language)
- Clone a specific voice or at least get good quality TTS output
- Run this on Colab or locally (preferably Python-based solutions)
If anyone has experience with multilingual TTS, voice cloning, or non-standard language models — I’d love to hear your recommendations or workflows that worked for you.

Yapaz said:

Hey everyone,
I’m working on a project aimed at generating text and converting it into speech for languages that are not commonly supported by most TTS tools (basically, anything outside English, Spanish, French, etc.).
The concept is simple: I generate text content based on a given theme or input (this part works fine with GPT models). The challenge is converting that generated text into audio — in a specific less-common language.
Here’s what I’ve tried so far:

• Coqui TTS:Tried multiple versions (0.22.0, 0.9.0, etc.). Ran into constant issues like:
  • No module named 'TTS.tts.train'
  • No module named 'TTS.trainer'
  • Missing or deprecated training modules
• gTTS (Google Text-to-Speech): Limited language support and the audio output quality isn’t good enough for this use case
• Voice cloning attempts: Faced compatibility issues, lack of documentation, or outdated repos when trying to build or fine-tune models on my datase

I’m open to any suggestions, libraries, or approaches that could help me:
- Train or fine-tune a TTS model on my own dataset (for a less-supported language)
- Clone a specific voice or at least get good quality TTS output
- Run this on Colab or locally (preferably Python-based solutions)
If anyone has experience with multilingual TTS, voice cloning, or non-standard language models — I’d love to hear your recommendations or workflows that worked for you.

Click to expand...

Have you checked out Mozilla's Common Voice project? It might offer some tools or community insights for training TTS models in less common languages.

Yapaz said:

Hey everyone,
I’m working on a project aimed at generating text and converting it into speech for languages that are not commonly supported by most TTS tools (basically, anything outside English, Spanish, French, etc.).
The concept is simple: I generate text content based on a given theme or input (this part works fine with GPT models). The challenge is converting that generated text into audio — in a specific less-common language.
Here’s what I’ve tried so far:

• Coqui TTS:Tried multiple versions (0.22.0, 0.9.0, etc.). Ran into constant issues like:
  • No module named 'TTS.tts.train'
  • No module named 'TTS.trainer'
  • Missing or deprecated training modules
• gTTS (Google Text-to-Speech): Limited language support and the audio output quality isn’t good enough for this use case
• Voice cloning attempts: Faced compatibility issues, lack of documentation, or outdated repos when trying to build or fine-tune models on my datase

I’m open to any suggestions, libraries, or approaches that could help me:
- Train or fine-tune a TTS model on my own dataset (for a less-supported language)
- Clone a specific voice or at least get good quality TTS output
- Run this on Colab or locally (preferably Python-based solutions)
If anyone has experience with multilingual TTS, voice cloning, or non-standard language models — I’d love to hear your recommendations or workflows that worked for you.

Click to expand...

1. Which language/s are we talking about here? And what is the usecase/quality you need? Do you need intination/prosody/sentiment? Do you need it in/near realtime for conversational use?

I may be able to guide you in the right direction, but a solution is very dependent on the required language/available data/research etc. There's a big difference between an 'easy' to replicate language and a 'hard' one (tonal, complex grapheme<->phoneme mapping, low resources/dataset/corpus) when it comes to how to approach this.

Depending on the usecase you can start by looking into:
- ToucanTTS (~7.000 languages)
- Training StyleTTS2 + maybe a new pl-bert model; F5-TTS; GPT-SoVITS etc.
- G2P (grapheme to phoneme) conversion tehniques/models
- What facebook did with mms-tts, though that's old stuff by today's standards

There are even new techniques/research such as https://arxiv.org/html/2402.18932v1 ; https://arxiv.org/html/2409.01217v1 etc.

2. Are you entirely sure nothing on the market already suits your needs? For example azure speech and other neural TTS solutions have very extensive language support (https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts) that can be fine-tuned if needed.