GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
akimbabananan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
akimbabananan
Model Tree
openai/whisper-base
Input Modalities
Output Modalities
Supported Functionality
Files.English Version: I completely understand why many find it frustrating that 99% of modern AI models, including this one, are trained for multilingual speech recognition. This creates an inconvenience for users who require specific languages for their target use cases. Here you will find a fully functional solution to this issue, including all the files required by both end-users and developers.
files directory.Эта модель представляет собой глубоко специализированную версию OpenAI Whisper Base, прошедшую процедуру жесткого полноразмерного дообучения (Full Fine-Tuning). Архитектура полностью перестроена и оптимизирована для высокоточного распознавания речи (ASR) в строго изолированной двуязычной среде (Русский и Английский языки).
Проект основан на концепции удаления избыточных языков из весов модели:
forced_decoder_ids = None) и 5000 шагов агрессивного обучения из модели были полностью вытеснены ассоциативные связи для 98 сторонних языков.eval_loss: 0.1929 был успешно достигнут на 4500-м шаге.gradient_checkpointing=True.gradient_accumulation_steps=4).learning_rate=5e-5 и warmup_steps=300.This model is a highly specialized version of OpenAI Whisper Base that has undergone a rigorous Full Fine-Tuning procedure. The architecture is fully optimized for high-precision Automatic Speech Recognition (ASR) within a strictly isolated bilingual environment (Russian and English).
The project is built on the concept of removing redundant languages from the model weights:
forced_decoder_ids = None) and applying 5000 steps of aggressive training, associative connections for 98 third-party languages were completely displaced from the model.eval_loss of 0.1929 was successfully achieved at step 4500.gradient_checkpointing=True activation method were applied.gradient_accumulation_steps=4).learning_rate=5e-5 and warmup_steps=300.Скрипт для быстрого тестирования модели напрямую из репозитория Hugging Face:
A script to quickly test the model directly from the Hugging Face repository:
python
import torchimport librosafrom transformers import WhisperProcessor, WhisperForConditionalGeneration# 1. Загрузка процессора и модели из репозиторияmodel_id = "akimbabananan/whisper-base-ru-en-only"device = "cuda" if torch.cuda.is_available() else "cpu"processor = WhisperProcessor.from_pretrained(model_id)model = WhisperForConditionalGeneration.from_pretrained(model_id).to(device)# 2. Подготовка аудио (укажите путь к вашему файлу)audio_path = "your_audio_file.wav"audio_input, sampling_rate = librosa.load(audio_path, sr=16000)# 3. Преобразование в признаки и генерация текстаinput_features = processor(audio_input, sampling_rate=sampling_rate, return_tensors="pt").input_features.to(device)forced_decoder_ids = processor.get_decoder_prompt_ids(language="russian", task="transcribe")with torch.no_grad():predicted_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids)# 4. Вывод результата (извлекаем [0], чтобы получить чистую строку)transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]print(f"Результат распознавания: {transcription}")