cstr / paraformer-zh-GGUF

Paraformer-zh — GGUF (ggml-quantised)

GGUF / ggml conversion of funasr/paraformer-zh for use with the paraformer backend in CrispStrobe/CrispASR .

Paraformer-zh is Alibaba's non-autoregressive ASR model (~220M params): a single forward pass through 50 SANM encoder blocks, a CIF (continuous integrate-and-fire) predictor, and 16 NAR decoder blocks produces the full transcript — no autoregressive token-by-token generation. Primarily Mandarin Chinese with English support. Character-level tokenizer (8404 vocab).

Architecture

Audio (16 kHz mono)
  → Kaldi fbank (80 mel, 25 ms / 10 ms)
  → LFR: stack 7, stride 6 → (T_lfr, 560)
  → CMVN (AddShift + Rescale, 560-dim)
  → SANMEncoder: 1 entry block (560→512) + 49 main blocks (512→512)
      each: LayerNorm → fused QKV → FSMN(k=11) + MHA(4 heads) → FFN(2048)
  → CifPredictorV2:
      Conv1d(512,512,k=3) → ReLU → Linear(512,1) → sigmoid
      → CIF accumulation (fire when alpha ≥ 1.0)
      → acoustic_embeds: (N_tokens, 512)
  → ParaformerSANMDecoder: 16 blocks
      each: norm1 → FFN → norm2 → FSMN(k=11) → norm3 → cross-attn(Q=dec, KV=enc)
  → decoders3: 1 post-processing block (FFN only)
  → after_norm → output_layer(512→8404) → argmax → characters

• Encoder reuses the same SANM block as Fun-ASR-Nano and SenseVoice
• Decoder block order is unusual: FFN → FSMN → cross-attn (not the more common self-attn → cross-attn → FFN)
• FSMN = depthwise conv (no Q/K/V self-attention in the decoder)
• Cross-attention uses a fused K+V projection from encoder output

Files

Quick Start

git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build --target crispasr-cli

# Chinese:
./build/bin/crispasr \
    --backend paraformer \
    -m /path/to/paraformer-zh-q4_k.gguf \
    -f chinese_audio.wav --no-prints
# → 正是因为存在绝对正义所以我们接受现实的相对正义...

# English:
./build/bin/crispasr \
    --backend paraformer \
    -m /path/to/paraformer-zh-q4_k.gguf \
    -f samples/jfk.wav --no-prints
# → and so my fellow americans ask not what your country can do for you ask what you can do for your country

# Or auto-download (resolves to Q4_K by default):
./build/bin/crispasr --backend paraformer -m auto -f audio.wav

Output format

The output is raw character-level text:

• Chinese : characters concatenated directly (no spaces) — standard for Chinese text
• English : word-level tokens with spaces inserted between consecutive English words; BPE continuation markers ( @@ ) handled internally
• No punctuation or casing — the model's character vocabulary has only lowercase English. Use --punc-model for punctuation restoration if needed.

Verification

All three quants (F16, Q4_K, Q8_0) produce byte-identical transcripts vs the upstream Python reference ( funasr.AutoModel.generate() ) on:

• Chinese (13 s asr_example.wav ): 66 characters, exact match
• English (11 s JFK samples/jfk.wav ): 26 tokens, exact match

The crispasr-diff paraformer harness captures 73 intermediate stages (mel features, 50 encoder layers, CIF alphas, acoustic embeds, 16 decoder layers, decoder output, generated text) for element-wise cosine-similarity comparison.

Converting from upstream

If you want to convert from the upstream PyTorch model yourself:

# Download upstream model
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download('funasr/paraformer-zh',
    local_dir='paraformer-zh-upstream',
    local_dir_use_symlinks=False)
"

# Convert to GGUF
python3 models/convert-paraformer-to-gguf.py \
    --input paraformer-zh-upstream \
    --output paraformer-zh-f16.gguf

# Quantize
./build/bin/crispasr-quantize paraformer-zh-f16.gguf paraformer-zh-q4_k.gguf q4_k
./build/bin/crispasr-quantize paraformer-zh-f16.gguf paraformer-zh-q8_0.gguf q8_0

Licence + attribution

Upstream funasr/paraformer-zh :

• Code (the funasr Python package): Apache-2.0.
• Model weights : FunASR Model License (Alibaba) — commercial use OK with attribution.

These GGUF files are a quantised / repackaged distribution of the upstream weights and inherit the FunASR Model License. Please attribute Alibaba / FunAudioLLM in downstream products.

If you use this model, please also cite the upstream FunASR work. See the upstream model card for the canonical citation.