Forwarded from Viacheslav
всім привіт. виклав свої тулзи для speech generation evaluation в паблік: https://github.com/balacoon/speech_gen_eval
і зробив leaderboard на основі: https://huggingface.co/spaces/balacoon/TTSLeaderboard.
більш детально тут: https://balacoon.com/blog/tts_leaderboard/
пишіть якщо хочете додати якусь модельку.
цікаво б ще було зробити testset для українських TTS систем, і додати в leaderboard.
це звісно не TTSArena, але повчально. в Leaderboard можна послухати семпли side-by-side тому доволі практично.
і зробив leaderboard на основі: https://huggingface.co/spaces/balacoon/TTSLeaderboard.
більш детально тут: https://balacoon.com/blog/tts_leaderboard/
пишіть якщо хочете додати якусь модельку.
цікаво б ще було зробити testset для українських TTS систем, і додати в leaderboard.
це звісно не TTSArena, але повчально. в Leaderboard можна послухати семпли side-by-side тому доволі практично.
GitHub
GitHub - balacoon/speech_gen_eval: Collection of tools to run objective evaluation of speech generation
Collection of tools to run objective evaluation of speech generation - balacoon/speech_gen_eval
🔥2
https://er.ucu.edu.ua/bitstreams/1faace86-5855-4737-af2c-d8559cbec52a/download
The author used my model to make evaluations
#asr #ai
The author used my model to make evaluations
#asr #ai
👍3
Please open Telegram to view this post
VIEW IN TELEGRAM
doing something
why it is so painful... https://huggingface.co/datasets/Helsinki-NLP/opus-100 #nlp
GitHub
GitHub - MicrosoftTranslator/NTREX: NTREX -- News Test References for MT Evaluation
NTREX -- News Test References for MT Evaluation. Contribute to MicrosoftTranslator/NTREX development by creating an account on GitHub.
https://predibase.com/blog/teaching-ai-to-write-gpu-code-a-deep-dive-into-reinforcement-fine-tuning
#ai #nlp
#ai #nlp
Predibase
Train AI to Write GPU Code via Reinforcement Fine-Tuning
In this post, we’ll discuss how we taught an AI model to convert PyTorch code into efficient Triton kernels using a reinforcement learning algorithm inspired by PPO called Group Relative Preference Optimization (GRPO).
doing something
https://huggingface.co/collections/jbochi/madlad-400-65491e6a78726cac9a4b84b7 #nlp
Tested this neural network.
Better than OPUS and NLLB for Ukrainian.
3B model requires about 12 GB of memory (float32).
Better than OPUS and NLLB for Ukrainian.
3B model requires about 12 GB of memory (float32).
doing something
Tested this neural network. Better than OPUS and NLLB for Ukrainian. 3B model requires about 12 GB of memory (float32).
Sometimes we need hacks to fix NN's issues
https://github.com/pemistahl/lingua-py helps in such cases
https://github.com/pemistahl/lingua-py helps in such cases