icefall-asr-ko-streaming-zipformer-72m

📖 한국어 README 보기 / Korean version

A Korean causal ASR built with the icefall recipe — 71.4M-parameter Zipformer transducer trained on KsponSpeech only.

• Training data: KsponSpeech, ~1,000 hours
• KsponSpeech eval CER (no-punctuation, CPU inference)
  • 7.68% — offline best (chunk=−1 PyTorch + KenLM, RTF 0.057)
  • 8.25% — streaming best (sherpa-onnx int8 chunk-64, RTF 0.038)
• Release: PyTorch .pt + sherpa-onnx int8 ONNX (chunk 16/32/64, ready to use out-of-the-box)

Model

Based on icefall Zipformer transducer.

• causal architecture (multichunk training: chunk-16 / 32 / 64 / -1)
• Korean syllable vocabulary (2,460 tokens)
• FP16 training
• Files included: PyTorch epoch-99.pt (3-way state-dict avg, 273 MB) + 9 sherpa-onnx int8 ONNX (chunk 16/32/64 × encoder/decoder/joiner). Ready to use out-of-the-box, no extra export step required

Performance (KsponSpeech eval, CER %, no-punctuation)

eval_clean 3,000 + eval_other 3,000 = 6,000 cuts. All measurements on the same set, beam=8, 3-way state-dict avg.

Decoding modes — choose by latency budget

Mode	chunk	clean	other	OVR	RTF	latency
chunk=−1 PyTorch + KenLM λ=0.10 (offline best)	−1	7.63	7.72	7.68	0.057 (CPU 2t)	utterance end
sherpa-onnx int8 chunk-64	64	8.28	8.22	8.25	0.038 (CPU 1t)	1.28 s
sherpa-onnx int8 chunk-32	32	8.45	8.33	8.39	0.039 (CPU 1t)	640 ms
sherpa-onnx int8 chunk-16	16	8.63	8.64	8.635	0.052 (CPU 1t)	320 ms

streaming numbers measured with 1x trailing padding (CHUNK * 0.04 s). RTF averaged over clean/other.

chunk=−1 (offline best 7.68%) is PyTorch-only: sherpa-onnx supports streaming chunks (16/32/64) only. To get the offline best CER, run epoch-99.pt directly via the icefall recipe — see "Option 1 — PyTorch (epoch-99.pt) with icefall recipe" below.

Quantization safety — fp32 vs int8 ⭐

Direct sweep on the same KsponSpeech 6,000-cut eval set, identical 1x padding, CPU 1-thread:

chunk	fp32 OVR	int8 OVR	Δ CER	fp32 RTF	int8 RTF	speedup
16	8.645	8.635	−0.010	0.078	0.052	1.50×
32	8.340	8.390	+0.050	0.060	0.039	1.54×
64	8.115	8.250	+0.135	0.057	0.038	1.50×

int8 quantization preserves CER (avg Δ +0.06%p, max +0.135) while running ~1.5× faster on CPU — this is the reason only .int8.onnx is shipped. If you need fp32, re-export from epoch-99.pt with the icefall recipe.

Direct comparison on the same eval set (KsponSpeech 6,000 cuts)

Model	params	clean	other	OVR
sibling 174M model	155.7M	7.18	7.14	7.16
This model (72M)	71.4M	7.63	7.72	7.68
Qwen3-ASR-1.7B (Alibaba)	1,700M	9.76	10.22	9.99
OpenAI Whisper-large-v2	1,550M	13.70	12.64	13.17

All models measured directly on the same KsponSpeech eval set. Whisper / Qwen run with default GPU FP16 transformers options; this model uses chunk=−1 PyTorch + KenLM rescore.

Decoding stack ablation

Stack	clean	other	OVR
epoch-40 single, no LM	8.11	8.06	8.08
3-way state-dict avg, no LM	7.81	7.84	7.83
3-way avg + KenLM λ=0.10 (the stack reported above)	7.63	7.72	7.68

3-way avg = state_dict average of best-valid + best-train + epoch-40.

Architecture

Encoder	Zipformer2 causal, 6 stacks
Encoder dim	192,256,384,512,384,256
Encoder layers	2,2,3,4,3,2
Feedforward dim	512,768,1024,1536,1024,768
Heads	4,4,4,8,4,4
Decoder	Stateless transducer decoder, dim 512
Joiner	dim 512
Vocabulary	Korean syllables, 2,460 tokens
Total params	71.4 M
Causal	Yes
Training chunk-size	16, 32, 64, −1 multichunk

Training

Data	KsponSpeech 1,000h (AIHub dataset 123)
Precision	FP16
Loss	Pruned-RNNT + CTC aux (scale 0.2)
Augmentation	SpecAugment (musan/RIR optional)
Optimizer	ScaledAdam
Total epochs	40
Released checkpoint	3-way state_dict avg of best-valid + best-train + epoch-40

Usage

Option 1 — PyTorch (epoch-99.pt) with icefall recipe

Load the weights with the icefall recipe (egs/librispeech/ASR/zipformer). chunk size, beam, KenLM, ONNX export, quantization etc. can be tuned freely.

chunk=−1 (offline best 7.68%) example:

# Place epoch-99.pt as epoch-99.pt in your icefall exp dir, then:
python zipformer/decode.py \
    --epoch 99 --avg 1 --use-averaged-model 0 \
    --causal 1 --chunk-size -1 --left-context-frames -1 \
    --decoding-method modified_beam_search --beam-size 8

Add KenLM rescore (λ=0.10) to reproduce 7.68%. See icefall docs for detailed options and KenLM usage.

Option 2 — sherpa-onnx streaming (recommended for CPU deployment)

# Install: pip install sherpa-onnx soundfile
import sherpa_onnx
import soundfile as sf

# Choose chunk: 16 (320ms latency) / 32 / 64
# Choose variant: ".int8" (included in release, ~1/2 size, almost no CER loss)
#                 "" (fp32, NOT included — re-export from epoch-99.pt if needed)
CHUNK = 16
VARIANT = ".int8"

recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
    encoder=f"encoder-epoch-99-avg-1-chunk-{CHUNK}-left-128{VARIANT}.onnx",
    decoder=f"decoder-epoch-99-avg-1-chunk-{CHUNK}-left-128{VARIANT}.onnx",
    joiner=f"joiner-epoch-99-avg-1-chunk-{CHUNK}-left-128{VARIANT}.onnx",
    tokens="tokens.txt",
    num_threads=2,
    sample_rate=16000,
    feature_dim=80,
    decoding_method="modified_beam_search",
    max_active_paths=8,
)

audio, sr = sf.read("speech.wav")
stream = recognizer.create_stream()
stream.accept_waveform(sr, audio)
# Add trailing silence so the last chunk is fully decoded
import numpy as np
stream.accept_waveform(16000, np.zeros(int(CHUNK * 0.04 * 16000), dtype=np.float32))
stream.input_finished()
while recognizer.is_ready(stream):
    recognizer.decode_stream(stream)
print(recognizer.get_result(stream))

Chunk size guide

chunk	latency	OVR	use case
16	320 ms	8.635	real-time UX, lowest interactive latency
32	640 ms	8.39	balanced
64	1.28 s	8.25	best streaming CER (only +0.57 vs offline 7.68)

Files

epoch-99.pt                                                  # PyTorch 3-way state_dict avg (273 MB)
tokens.txt                                                   # syllable vocab 2,460
README.md / README_ko.md
encoder-epoch-99-avg-1-chunk-{16,32,64}-left-128.int8.onnx   # int8 encoder
decoder-epoch-99-avg-1-chunk-{16,32,64}-left-128.int8.onnx
joiner-epoch-99-avg-1-chunk-{16,32,64}-left-128.int8.onnx

Sibling

→ icefall-asr-ko-streaming-zipformer-174m — 155.7M params, KSP + AIHub 6,500h.

License

Apache 2.0.

Citation

@misc{ko-zipformer-72m,
  title  = {icefall-asr-ko-streaming-zipformer-72m: Korean ASR (KsponSpeech)},
  author = {Ken (kangkyu)},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/kangkyu/icefall-asr-ko-streaming-zipformer-72m}}
}

Built on:

• icefall
• k2
• KsponSpeech