docs: improve speech-to-text discovery

Author: ai-babai

README.md

@@ -2,16 +2,21 @@
 
 **English** · [Русский](README.ru.md)
 
-Fast, offline speech-to-text for Russian, Kazakh, Kyrgyz, and Uzbek on Apple
-Silicon. This is an independent native [MLX](https://github.com/ml-explore/mlx)
-port of [GigaAM](https://github.com/salute-developers/GigaAM), based on the
-official [GigaAM-Multilingual model](https://huggingface.co/ai-sage/GigaAM-Multilingual).
+Fast, private, offline speech-to-text and audio transcription for Russian,
+Kazakh, Kyrgyz, and Uzbek on Apple Silicon Macs. Transcribe meetings,
+interviews, lectures, podcasts, voice messages, and video locally to TXT, JSON,
+SRT, or VTT. This is an independent native
+[MLX](https://github.com/ml-explore/mlx) port of
+[GigaAM](https://github.com/salute-developers/GigaAM), based on the official
+[GigaAM-Multilingual model](https://huggingface.co/ai-sage/GigaAM-Multilingual).
 
 [PyPI](https://pypi.org/project/gigaam-multilingual-mlx/) ·
 [Hugging Face models](https://huggingface.co/collections/ai-babai/gigaam-multilingual-mlx-6a57ef36e44d1a0c4eb49276) ·
 [Latest release](https://github.com/ai-babai/gigaam-multilingual-mlx/releases/latest) ·
 [Full benchmark](docs/benchmark-multilingual-v1.md)
 
+## Speech recognition quality by language
+
 ![GigaAM MLX, Whisper, and Parakeet model decision matrix](https://raw.githubusercontent.com/ai-babai/gigaam-multilingual-mlx/main/docs/benchmark-multilingual-v1.png?v=0.1.5)
 
 The chart compares WER, five-minute transcription time, peak process memory, and

README.ru.md

@@ -2,8 +2,11 @@
 
 [English](README.md) · **Русский**
 
-Быстрое локальное распознавание русской, казахской, кыргызской и узбекской речи
-на Apple Silicon. Это независимый нативный порт
+Быстрое локальное распознавание речи в текст и транскрибация аудио на Mac
+с Apple Silicon — для русского, казахского, кыргызского (киргизского) и узбекского
+языков. GigaAM MLX локально расшифровывает встречи, интервью, лекции, подкасты,
+голосовые сообщения и аудиодорожки видео в TXT, JSON, SRT или VTT. Это
+независимый нативный порт
 [GigaAM](https://github.com/salute-developers/GigaAM) на
 [MLX](https://github.com/ml-explore/mlx), основанный на официальной модели
 [GigaAM-Multilingual](https://huggingface.co/ai-sage/GigaAM-Multilingual).
@@ -13,6 +16,8 @@
 [Последний релиз](https://github.com/ai-babai/gigaam-multilingual-mlx/releases/latest) ·
 [Полный бенчмарк](docs/benchmark-multilingual-v1.md)
 
+## Качество распознавания по языкам
+
 ![Сравнение GigaAM MLX, Whisper и Parakeet](https://raw.githubusercontent.com/ai-babai/gigaam-multilingual-mlx/main/docs/benchmark-multilingual-v1.png?v=0.1.5)
 
 На графике сравниваются WER, время расшифровки пятиминутного WAV, пиковая память

tools/prepare_hf_release.py

@@ -165,7 +165,18 @@ def _card(
         if variant == "int8"
         else "Select this variant explicitly with `--variant`."
     )
-    yaml_tags = ["mlx", "apple-silicon", "gigaam", "ctc", "automatic-speech-recognition"]
+    yaml_tags = [
+        "mlx",
+        "apple-silicon",
+        "gigaam",
+        "ctc",
+        "automatic-speech-recognition",
+        "speech-recognition",
+        "transcription",
+        "offline-asr",
+        "speech-to-text",
+        "macos",
+    ]
     if quantization:
         yaml_tags.extend(["quantized", precision])
     lines = [
@@ -187,7 +198,11 @@ def _card(
         "",
         f"# {metadata['title']}",
         "",
-        "An independent Apple MLX port of the official "
+        "Native offline speech recognition and audio transcription for Russian, Kazakh, "
+        "Kyrgyz, and Uzbek on Apple Silicon with MLX. Core speech-to-text languages: "
+        "Russian (русский), Kazakh (қазақша), Kyrgyz (кыргызча), and Uzbek (o‘zbekcha).",
+        "",
+        "This is an independent Apple MLX port of the official "
         f"[GigaAM repository]({UPSTREAM_REPOSITORY}) and the "
         f"[GigaAM-Multilingual source model]({UPSTREAM_MODEL}). It is not an official "
         "GigaChat/GigaAM release.",