Somali Voice Models

Somali is spoken by about 25 million people, yet there is still no production-ready support from commercial services like ElevenLabs or Cartesia.

Results from hands-on experiments are as follows.

• MMS-TTS (facebook/mms-tts-som): A working baseline, but the quality leaves something to be desired.
• Fish Speech V1.5 LoRA: Showed potential, but the pronunciation wasn't clean enough.
• XTTS V4: The best results so far, trained up to 235K steps on about 300 hours of Somali speech data.

The biggest issue is that the tokenizer has no [so] token. Since Somali uses the Latin script, the [en] token was used as a workaround.

TTS pronunciation keeps improving, but the harder problem is the LLM layer. Most models have seen very little Somali text, resulting in weak comprehension and unnatural response generation. Whisper also has low transcription accuracy for Somali.

Ultimately, this is a request for people to share what actually works well for low-resource African languages like Somali, Amharic, and Tigrinya.