You can deploy this model and start earning money today!

IconicAI/
sherpa-ncnn-zipformer-libri-giga-70M-2023-04-04

Name: sherpa-ncnn-zipformer-libri-giga-70M-2023-04-04
Availability: InStock

automatic-speech-recognition

apache-2.0

Views

No views yet

icefall-libri-giga-pruned-transducer-stateless7-streaming-2023-04-04 — NCNN export (streaming Zipformer, English)

This repo provides an NCNN export of the icefall streaming Zipformer model trained on LibriSpeech + GigaSpeech (2023-04-04).
It is ready for sherpa-ncnn (Android and desktop).

Files included

encoder_jit_trace-pnnx.ncnn.param
encoder_jit_trace-pnnx.ncnn.bin
decoder_jit_trace-pnnx.ncnn.param
decoder_jit_trace-pnnx.ncnn.bin
joiner_jit_trace-pnnx.ncnn.param
joiner_jit_trace-pnnx.ncnn.bin
tokens.txt (500 tokens; blank id = 0)

The encoder .param already contains the required SherpaMetaData layer. Files are UTF-8 without BOM to match NCNN’s plain-text parser.

Model details (export)

Architecture: Streaming Zipformer Transducer (RNN-T)
Context size: 2
Encoder stacks / layers: 2,4,3,2,4
Encoder dims: all 384
Attention dims: all 192
Downsampling factors: 1,2,4,8,2
CNN kernels: all 31
Decoder dim / Joiner dim: 512 / 512
Feature extractor: 80-dim log-Mel at 16 kHz
Streaming chunking: decode_chunk_len=32, num_left_chunks=4, pad_len=7
Export toolchain:
- PyTorch 1.13.1 (CPU)
- pnnx (Fangjun’s csukuangfj/ncnn fork, MSVC build on Windows)
- Post-step: insert SherpaMetaData into the encoder .param
- Filenames follow the hyphen-pnnx convention for drop-in compatibility

How to use (sherpa-ncnn)

Android (JNI)

Copy the six NCNN files and tokens.txt into your app’s assets/ (or load from files), then construct the recognizer with those paths. Ensure your NCNN build includes ops used by modern Zipformers (e.g., CumulativeSum). If Vulkan isn’t present, useGPU falls back to CPU automatically.

Desktop quick start (CLI)

If you use the sherpa-ncnn CLI, point each component explicitly (paths adjusted):

bash

sherpa-ncnn \
  --encoder-param  encoder_jit_trace-pnnx.ncnn.param \
  --encoder-bin    encoder_jit_trace-pnnx.ncnn.bin   \
  --decoder-param  decoder_jit_trace-pnnx.ncnn.param \
  --decoder-bin    decoder_jit_trace-pnnx.ncnn.bin   \
  --joiner-param   joiner_jit_trace-pnnx.ncnn.param  \
  --joiner-bin     joiner_jit_trace-pnnx.ncnn.bin    \
  --tokens         tokens.txt                        \
  --decoding-method greedy_search                    \
  --num-threads 4

IconicAI/sherpa-ncnn-zipformer-libri-giga-70M-2023-04-04