cstr/mega-asr-GGUF

Mega-ASR 1.7B — GGUF

GGUF / ggml conversions of zhifeixie/Mega-ASR for use with CrispStrobe/CrispASR.

Mega-ASR is a robust ASR adaptation of Qwen3-ASR-1.7B. The upstream release consists of:

• Qwen3-ASR-1.7B base model
• mega-asr-merged/adapter_model.safetensors, a LoRA adapter trained for degraded / in-the-wild audio
• audio_quality_router/best_acc_model.safetensors, a lightweight router that decides whether to enable the LoRA

This GGUF release uses the always-on robust path: the Mega-ASR LoRA is merged into the Qwen3-ASR-1.7B weights offline, then converted to a normal Qwen3-ASR GGUF. CrispASR runs it through the standard qwen3 backend; no Python, PEFT, safetensors, or runtime LoRA switching is needed.

Files

File	Size	Notes
mega-asr-1.7b-f16.gguf	4.4 GB	F16 merged-LoRA model
mega-asr-1.7b-q4_k.gguf	1.3 GB	Q4_K, recommended default
mega-asr-router.gguf	2.8 MB	Converted upstream audio-quality router weights; auxiliary artifact, not used by current CrispASR inference

Quick Start

Build CrispASR:

git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target crispasr

Download the recommended quant:

huggingface-cli download cstr/mega-asr-GGUF \
    mega-asr-1.7b-q4_k.gguf --local-dir .

Transcribe:

./build/bin/crispasr \
    --backend mega-asr \
    -m mega-asr-1.7b-q4_k.gguf \
    your-audio.wav

Older CrispASR builds can use the same file with the canonical backend name:

./build/bin/crispasr \
    --backend qwen3 \
    -m mega-asr-1.7b-q4_k.gguf \
    your-audio.wav

Auto-Download

Current CrispASR main registers mega-asr as an alias for the Qwen3-ASR runtime:

./build/bin/crispasr --backend mega-asr -m auto your-audio.wav

This downloads mega-asr-1.7b-q4_k.gguf and runs it with the existing Qwen3-ASR backend.

Validation

Smoke test on samples/jfk.wav, CPU-only VPS build:

./build/bin/crispasr \
    --backend mega-asr \
    -m mega-asr-1.7b-q4_k.gguf \
    --no-timestamps \
    samples/jfk.wav

Output:

And so, my fellow Americans, ask not what your country can do for you. Ask what you can do for your country.

The runtime reports the expected Mega-ASR / Qwen3-ASR-1.7B geometry:

audio 24 layers, llm 28 layers, vocab 151936

Router Status

The upstream Mega-ASR router chooses between the clean base model and the robust LoRA path. This GGUF release currently ships the practical always-on robust model.

The small mega-asr-router.gguf file is included so the router weights are available in GGUF form, but CrispASR does not yet use it to switch between base and LoRA models at runtime. Dynamic routing would require either loading both Qwen3-ASR-1.7B variants or adding runtime LoRA switching.

The router's frontend is a LogMelSpectrogram using 16 kHz audio, 80 Slaney mel bins, n_fft=400, hop_length=160, win_length=400, log10(clamp(mel, 1e-10)), then (x + 4) / 4. CrispASR already has the reusable pieces for this in src/core/mel.h (build_slaney_fb + core_mel::compute()), so if dynamic routing is added later it should reuse that shared mel path rather than adding another spectrogram implementation.

For degraded audio, the always-on merged model is the intended useful path. For consistently clean audio, use the regular Qwen3-ASR models.

How this was made

1. Downloaded upstream zhifeixie/Mega-ASR.
2. Merged mega-asr-merged/adapter_model.safetensors into Qwen3-ASR-1.7B offline.
3. Converted the merged Hugging Face checkpoint with CrispASR's existing Qwen3-ASR converter.
4. Quantised the F16 GGUF with CrispASR's crispasr-quantize.

Upstream

• Model weights: zhifeixie/Mega-ASR
• Code: xzf-thu/Mega-ASR
• Paper: Mega-ASR: Towards In-the-Wild^2 Speech Recognition via Scaling Up Real-world Acoustic Simulation
• Base model: Qwen/Qwen3-ASR-1.7B

License

Apache-2.0, inherited from the upstream Mega-ASR project and Qwen3-ASR base model.

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/GGML). 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.