cstr/qwen3-asr-1.7b-GGUF

Qwen3-ASR-1.7B — GGUF

GGUF quantizations of Qwen/Qwen3-ASR-1.7B for use with CrispASR.

Qwen3-ASR-1.7B is a multilingual speech recognition model supporting 30 languages + 22 Chinese dialects, with support for speech, singing voice, and songs with background music. Audio inputs up to 5 minutes.

Files

File	Quant	Size	Notes
qwen3-asr-1.7b-q4_k.gguf	Q4_K	~1.5 GB	Recommended for most use cases (audio tower kept at Q8_0 — see rebake note)
qwen3-asr-1.7b-q8_0.gguf	Q8_0	~2.5 GB	Higher quality, more memory
qwen3-asr-1.7b-f16.gguf	F16	~4.7 GB	Full precision (needs ~5 GB RAM)

2026-07 rebake — audio tower now Q8_0 (fixes empty / repeated transcripts)

qwen3-asr-1.7b-q4_k.gguf was re-baked with the 24-layer audio encoder kept at Q8_0 (previously Q4_K like the LLM body; ~0.15 GB larger). The earlier Q4_K export could emit an empty transcript (or, on longer audio, a repeated phrase) while still reporting success (CrispASR #240) — the same sub-8-bit encoder drift diagnosed in #218. Encoder-output cosine vs the F16 reference (jfk): 0.9632 → 0.9989 min, 0.9913 → 0.9998 mean — i.e. the tower is back to near-lossless and decode no longer degenerates. The Q8_0 and F16 files were never affected.

Usage with CrispASR

# Auto-download and transcribe (Q4_K)
crispasr -m qwen3-1.7b -f audio.wav

# Or specify the model path directly
crispasr -m qwen3-asr-1.7b-q4_k.gguf -f audio.wav --backend qwen3

# With language hint
crispasr -m qwen3-1.7b -f audio.wav -l ja

# Translation mode
crispasr -m qwen3-1.7b -f audio.wav --translate --target-lang en

Performance

On JFK speech sample (11s, English):

• Q4_K: 0.2x realtime on CPU (4 threads)
• Perfect transcription accuracy

Conversion

Converted from the non-HF variant (Qwen/Qwen3-ASR-1.7B, not the -hf variant):

python models/convert-qwen3-asr-to-gguf.py \
    --input Qwen/Qwen3-ASR-1.7B \
    --output qwen3-asr-1.7b-f16.gguf
crispasr-quantize qwen3-asr-1.7b-f16.gguf qwen3-asr-1.7b-q4_k.gguf q4_k

Note: The converter also supports the -hf variant (Qwen/Qwen3-ASR-1.7B-hf) which uses a different tensor naming convention. Both produce identical results.

Architecture

• Audio encoder: 24-layer Whisper-style encoder (d=1024, 16 heads, GELU)
• Projector: 2-layer MLP (1024 → GELU → 2048)
• LLM decoder: 28-layer Qwen3 (d=2048, 16Q/8KV heads, head_dim=128, SwiGLU)
• Vocab: 151,936 tokens (GPT-2 BPE)
• RoPE theta: 1,000,000

License

Apache 2.0

Provenance and EU AI Act Art. 53 note

• Upstream model: Qwen/Qwen3-ASR-1.7B — published by Qwen.
• 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 data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.