Belle-whisper-large-v3-zh is a Chinese-enhanced Whisper large-v3 model fine-tuned by the BELLE team. This repository provides the Ascend NPU adaptation for inference on Huawei Ascend hardware (Ascend 910B2).
Validation Status: PASS | Encoder Precision PASS | Decoder Precision PASS | E2E Transcription PASS
| Item | Details |
|---|---|
| Model | Belle-whisper-large-v3-zh |
| Architecture | WhisperForConditionalGeneration (Encoder-Decoder) |
| Original Source | Xorbits/Belle-whisper-large-v3-zh (ModelScope) |
| Parameters | ~1.5B |
| d_model | 1280 |
| Encoder Layers | 32 |
| Decoder Layers | 32 |
| Attention Heads | 20 |
| FFN Dimension | 5120 |
| Vocabulary | 51866 |
| Mel Bins | 128 |
| Precision | FP32 |
| Language | Chinese (zh) |
| Task | Automatic Speech Recognition (ASR) |
| Component | Version |
|---|---|
| NPU | Ascend 910B2 (2x, 64GB HBM each) |
| CANN | 8.5.1 |
| PyTorch | 2.9.0 |
| torch_npu | 2.9.0.post1+gitee7ba04 |
| transformers | >= 4.34.0 |
| Python | 3.11 |
pip install modelscope
modelscope download --model Xorbits/Belle-whisper-large-v3-zh \
--local_dir ./models/Belle-whisper-large-v3-zh \
--exclude "pytorch_model.bin"pip install torch torch_npu transformers librosa soundfile# Single audio file transcription
python inference.py --audio /path/to/audio.wav
# Benchmark mode
python inference.py --benchmark --num-runs 50python evaluate.py --num-runs 30 --warmup 5Tested on Ascend 910B2, FP32 precision, single NPU (npu:0).
| Metric | Value |
|---|---|
| Avg Latency | 0.2089 s |
| P50 Latency | 0.2085 s |
| P90 Latency | 0.2115 s |
| P99 Latency | 0.2156 s |
| Std Dev | 0.0023 s |
| Throughput | 9.58 tokens/s |
| Peak Memory | 6526 MB |
| Audio Duration | Input Frames | Avg Latency | Output Tokens |
|---|---|---|---|
| 3s | 3000 | 0.216 s | 2.0 |
| 5s | 3000 | 0.223 s | 2.0 |
| 10s | 3000 | 0.212 s | 2.0 |
| 15s | 3000 | 0.340 s | 5.0 |
| 30s | 3000 | 0.817 s | 15.0 |
Note: Synthetic tone audio produces minimal output tokens. Real speech would produce more tokens proportionally to audio content. Encoder runs in ~50ms (fixed cost), decoder generates ~10 tok/s.
NPU (FP32) outputs compared against CPU (FP32) reference on identical inputs.
| Metric | Value | Threshold | Status |
|---|---|---|---|
| Cosine Similarity | 0.9998005 | > 0.999 | PASS |
| SNR | 33.95 dB | > 30 dB | PASS |
| Max Absolute Error | 1.978 | - | - |
| Mean Absolute Error | 0.007 | - | - |
| Relative Error | 1.63% | - | - |
| Frobenius Relative | 0.0201 | - | - |
| Metric | Value | Threshold | Status |
|---|---|---|---|
| Cosine Similarity | 0.9999263 | > 0.999 | PASS |
| SNR | 29.95 dB | > 20 dB | PASS |
| Top-1 Overlap | 100% | - | - |
| Top-5 Overlap | 100% | >= 80% | PASS |
| Top-10 Overlap | 100% | - | - |
| Top-100 Overlap | 99% | - | - |
| Relative Error | 3.33% | < 5% | PASS |
| Metric | Value | Threshold | Status |
|---|---|---|---|
| Token Match Accuracy | 100% | > 90% | PASS |
| Exact Transcript Match | True | - | PASS |
| CPU Output | 嗯嗯 | - | - |
| NPU Output | 嗯嗯 | - | - |
All three validation stages pass. The encoder hidden states achieve cosine similarity > 0.9998. The decoder logits achieve cosine similarity > 0.9999. End-to-end transcription produces bit-identical results.
Belle-whisper-large-v3-zh/
├── README.md # This document
├── inference.py # NPU inference script (single audio / benchmark)
├── evaluate.py # Precision & performance evaluation suite
├── eval_report.json # Full evaluation results (JSON)
└── test_audio/ # Test audio generation script
└── generate_test.pyThe Whisper encoder-decoder pipeline on NPU:
This model adaptation is based on Xorbits/Belle-whisper-large-v3-zh on ModelScope. Please refer to the original model for license terms.
#NPU #Ascend #Whisper #ASR #Speech-Recognition #Hardware-NPU