Qwen3 Asr 1.7b Ja Anime GGUF

--- license: apache-2.0 language: - ja - en - zh tags: - asr - speech-recognition - gguf - qwen3 base_model: jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf pipeline_tag: automatic-speech-recognition --- # Qwen3-ASR-1.7B-JA-Anime-Galgame — GGUF GGUF quantizations of [jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf](https://huggingface.co/jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf) for use with [CrispASR](https://github.com/CrispStrobe/CrispASR). This is a Japanese anime/galgame fine-tune of Qwen3-ASR-1.7B. It supports 30+ languages with enhanced performance on Japanese anime/game speech. ## Files | File | Quant | Size | Notes | |------|-------|------|-------| | `qwen3-asr-1.7b-ja-anime-q4_k.gguf` | Q4_K | ~1.3 GB | Recommended for most use cases | | `qwen3-asr-1.7b-ja-anime-q8_0.gguf` | Q8_0 | ~2.5 GB | Higher quality, more memory | ## Usage with CrispASR ```bash # Auto-download and transcribe crispasr -m qwen3-ja-anime -f audio.wav # Or specify the model path directly crispasr -m qwen3-asr-1.7b-ja-anime-q4_k.gguf -f audio.wav --backend qwen3 ``` ## Conversion Converted from safetensors using: ```bash python models/convert-qwen3-asr-to-gguf.py \ --input jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf \ --output qwen3-asr-1.7b-ja-anime-f16.gguf crispasr-quantize qwen3-asr-1.7b-ja-anime-f16.gguf qwen3-asr-1.7b-ja-anime-q4_k.gguf q4_k ``` ## Architecture Same as Qwen3-ASR-1.7B: - Audio encoder: 24-layer Whisper-style encoder (d=1024, 16 heads) - Projector: 2-layer MLP (1024 → GELU → 2048) - LLM decoder: 28-layer Qwen3 (d=2048, 16Q/8KV heads, SwiGLU) - Vocab: 151,936 tokens (GPT-2 BPE) ## License Apache 2.0 (following the base model license) ## Provenance and EU AI Act Art. 53 note - **Upstream model:** [jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf](https://huggingface.co/jaykwok/Qwen3-ASR-1.7B-JA-Anime-Galgame-hf) — published by `jaykwok`. - **Upstream licence:** `apache-2.0`. This repository redistributes under the same terms; it grants no rights the upstream licence does not. - **What was done here:** format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs. - **Training data:** documented — where it is documented at all — by the upstream provider; see the upstream model card. No training...