cstr/firered-asr2-aed-GGUF

FireRedASR2-AED -- GGUF

GGUF conversions and quantisations of FireRedTeam/FireRedASR2-AED for use with CrispStrobe/CrispASR.

Available variants

File	Quant	Size	Notes
firered-asr2-aed.gguf	F16	2.3 GB	Full precision
firered-asr2-aed-q8_0.gguf	Q8_0	1.4 GB	High quality
firered-asr2-aed-q4_k.gguf	Q4_K	919 MB	Best size/quality tradeoff

All variants produce identical transcription on test audio.

Model details

• Architecture: Conformer encoder (16L, d=1280, 20 heads, relative positional encoding with pos_bias_u/v, macaron FFN, depthwise separable conv k=33) + CTC head
• Parameters: 1.1B
• Languages: Mandarin Chinese, English, 20+ Chinese dialects
• License: Apache 2.0
• CER: 3.05% (Mandarin average, per paper)
• Encoder: Hybrid ggml/CPU — ggml for matmuls, CPU for relative position attention scoring

Usage with CrispASR

git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8

# Auto-detect backend from GGUF
./build/bin/crispasr -m firered-asr2-aed-q4_k.gguf -f audio.wav

# Explicit backend
./build/bin/crispasr --backend firered-asr -m firered-asr2-aed-q4_k.gguf -f audio.wav -osrt

Note: Output is in UPPERCASE (the model was trained with uppercase English text). CTC decoding is used; beam search decoder not yet implemented.

Conversion

python models/convert-firered-asr-to-gguf.py --input FireRedTeam/FireRedASR2-AED --output firered-asr2-aed.gguf
crispasr-quantize firered-asr2-aed.gguf firered-asr2-aed-q4_k.gguf q4_k