Skip to content

Latest commit

 

History

History
64 lines (44 loc) · 2.5 KB

File metadata and controls

64 lines (44 loc) · 2.5 KB

API 速查

以下为稳定的公共接口。完整签名以源码 docstring 为准。

顶层导出(sepkit

名称 说明
stft(x, n_fft=512, hop_length=None, window="hann", center=True) 短时傅里叶变换
istft(spec, n_fft=None, hop_length=None, window="hann", center=True, length=None) 逆变换
ideal_binary_mask(sources) 理想二值掩蔽 IBM
ideal_ratio_mask(sources) 理想比值掩蔽 IRM
wiener_mask(sources) 维纳型(功率比)掩蔽
phase_sensitive_mask(sources, mixture=None, clip=(0,1)) 相位敏感掩蔽 PSM
complex_ratio_mask(sources, mixture=None) 复数比值掩蔽 cIRM
reconstruct_waveforms(mixture, mask, ...) 掩蔽作用 + 逆 STFT
si_sdr(estimate, reference) SI-SDR(dB)
si_sdr_improvement(estimate, reference, mixture) SI-SDRi(dB)
snr(estimate, reference) 尺度相关 SNR(dB)
bss_eval_sources(estimate, references) 返回 BSSMetrics(sdr, sir, sar)
evaluate_separation(estimates, references, mixture=None) 置换感知整体评估
resolve_permutation(estimates, references, metric, maximize=True) 最优置换求解
OracleTFMaskSeparator(references, mask="irm", ...) oracle 时频掩蔽分离器
mix_sources(sources, snr_db=None) 按目标 SNR 混合,返回 Mixture

子模块

sepkit.dsp

stft, istft, get_window, check_cola, frame_signal, overlap_add, num_frames

sepkit.masking

上述掩蔽函数 + apply_mask, reconstruct_waveforms, REAL_MASKS, available_masks

sepkit.pit

resolve_permutation, reorder, pairwise_matrix, PITResult, hungarian, all_permutations, best_permutation_bruteforce, score_all_permutations

sepkit.metrics

si_sdr, si_sdr_improvement, snr, bss_eval_sources, BSSMetrics, evaluate_separation, EvaluationReport

sepkit.simulation

tone, harmonic, chirp, white_noise, mix_sources, Mixture

sepkit.backend(需 sepkit[torch]

torch_si_sdr(estimate, reference), pit_sisdr_loss(estimates, references)

sepkit.io

read_wav(path) -> (data, sample_rate), write_wav(path, data, sample_rate=16000)

返回类型

  • Mixture(mixture, sources)
  • SeparationResult(sources, sample_rate) —— 属性 n_sources
  • PITResult(permutation, score, pairwise)
  • BSSMetrics(sdr, sir, sar)
  • EvaluationReport(permutation, si_sdr, si_sdri, sdr, sir, sar, extra) —— 属性 mean_si_sdr, mean_si_sdri