Urdu Model
As-salamu alaykum / Dear Rhasspy Team,
My name is Muhammad Noman and I am from Pakistan. I am a big admirer of the Piper TTS project — it is the ONLY offline TTS engine that runs smoothly on CPU and mobile devices. For Urdu speakers, this is a dream tool.
I have been working for several months to build a high-quality Urdu TTS model using Piper. I do not have a heavy PC — I have been training on Google Colab (free tier) by rotating Gmail accounts to work around session limits. Despite all these difficulties, I kept going because I believe in this project.
MY SETUP:
- Base checkpoint: https://huggingface.co/IhorShevchuk/piper-voice-ur-fasih
- Fine-tuned with thousands of my own voice recordings
- Training done on Google Colab T4 GPU
- Current checkpoint: epoch=3399-step=6948
- Live demo of my current model: https://huggingface.co/spaces/nzgnzg73/Urdu_TTS_Realistic
PROBLEM 1 — AUDIO DISTORTION (آواز پھٹ جانا):
My original recordings were completely clean and clear. But the generated audio sounds distorted and broken. I suspect this is a sample rate mismatch — my source audio was 16kHz but Piper requires 22050Hz. However, even after attempting to fix this, the distortion remains. Please listen to my demo space to hear the exact problem.
PROBLEM 2 — REPEATED WORD MISPRONUNCIATION (لفظ تھتراہٹ):
When a word like 'پاکستان' (Pakistan) appears once in a sentence, it is read correctly. But if the same word appears 2 or 3 times in the same text, it is mispronounced on the second or third occurrence. This seems to be an eSpeak Urdu phonemizer limitation — the Urdu eSpeak support appears to be quite weak for handling repeated or complex Urdu words.
MY REQUESTS:
Request 1 — Is there a better phonemizer for Urdu?
eSpeak's Urdu support is limited. Is there any alternative phonemizer that works better for Urdu? Or is there a custom phoneme dictionary approach that could help?
Request 2 — Mixed language support (Urdu + Hindi + English):
In Pakistan, people naturally mix Urdu, Hindi words, and English words in daily speech. For example:
UR_001|پاکستان زندہ باد
HI_001|नमस्ते کیسے ہو آپ
EN_001|Hello how are you today
MIX_001|Hello بھائی کیسے ہو آج
Is it possible to train a single Piper model that handles all three languages naturally? If yes, what is the recommended data format and training approach?
Request 3 — Emotional voice support:
I want to add Natural, Happy, Sad, and Angry voice styles. I am ready to record and provide all this audio data myself. I have already prepared a large amount of audio. Is there a recommended way to train multi-style/emotion models in Piper?
MY OFFER:
I am willing to contribute my Urdu voice dataset to the Piper project — including natural, happy, angry, and sad voice recordings — if it can help improve Urdu support for everyone. There are hundreds of millions of Urdu speakers worldwide and right now there is NO good offline Urdu TTS model available. Piper is the only hope.
I have very limited resources (no heavy PC, only free Colab) but I have passion and dedication. Any guidance from your team would mean the world to me and to the entire Urdu-speaking community.
Thank you so much for this amazing project.
Best regards,
Muhammad Noman
Pakistan
Project: NZG73
Demo: https://huggingface.co/spaces/nzgnzg73/Urdu_TTS_Realistic