Support computing features for whisper

CMakeLists.txt

@@ -7,7 +7,7 @@ project(kaldifeat)
 # remember to change the version in
 # scripts/conda/kaldifeat/meta.yaml
 # scripts/conda-cpu/kaldifeat/meta.yaml
-set(kaldifeat_VERSION "1.25.1")
+set(kaldifeat_VERSION "1.25.2")
 
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/lib")
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/lib")

kaldifeat/csrc/CMakeLists.txt

@@ -10,6 +10,7 @@ set(kaldifeat_srcs
   matrix-functions.cc
   mel-computations.cc
   online-feature.cc
+  whisper-fbank.cc
 )
 
 add_library(kaldifeat_core ${kaldifeat_srcs})

kaldifeat/csrc/CPPLINT.cfg

@@ -0,0 +1 @@
+exclude_files=whisper-mel-bank.h

kaldifeat/csrc/feature-fbank.cc

@@ -65,7 +65,7 @@ torch::Tensor FbankComputer::Compute(torch::Tensor signal_raw_log_energy,
   // note spectrum is in magnitude, not power, because of `abs()`
 #if defined(KALDIFEAT_HAS_FFT_NAMESPACE)
   // signal_frame shape: [x, 512]
-  // spectrum shape [x, 257
+  // spectrum shape [x, 257]
   torch::Tensor spectrum = torch::fft::rfft(signal_frame).abs();
 #else
   // signal_frame shape [x, 512]

kaldifeat/csrc/feature-window.cc

@@ -29,6 +29,13 @@ FeatureWindowFunction::FeatureWindowFunction(const FrameExtractionOptions &opts,
   float *window_data = window.data_ptr<float>();
 
   double a = M_2PI / (frame_length - 1);
+
+  if (opts.window_type == "hann") {
+    // see https://pytorch.org/docs/stable/generated/torch.hann_window.html
+    // We assume periodic is true
+    a = M_2PI / frame_length;
+  }
+
   for (int32_t i = 0; i < frame_length; i++) {
     double i_fl = static_cast<double>(i);
     if (opts.window_type == "hanning") {
@@ -39,6 +46,8 @@ FeatureWindowFunction::FeatureWindowFunction(const FrameExtractionOptions &opts,
       window_data[i] = sin(0.5 * a * i_fl);
     } else if (opts.window_type == "hamming") {
       window_data[i] = 0.54 - 0.46 * cos(a * i_fl);
+    } else if (opts.window_type == "hann") {
+      window_data[i] = 0.50 - 0.50 * cos(a * i_fl);
     } else if (opts.window_type ==
                "povey") {  // like hamming but goes to zero at edges.
       window_data[i] = pow(0.5 - 0.5 * cos(a * i_fl), 0.85);

kaldifeat/csrc/generate-whisper-melbank.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import librosa
+import numpy as np
+
+
+def main():
+    m = librosa.filters.mel(sr=16000, n_fft=400, n_mels=80)
+    assert m.shape == (80, 201)
+    s = "// Auto-generated. Do NOT edit!\n\n"
+    s += "// Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)\n\n"
+    s += "\n"
+    s += "#ifndef KALDIFEAT_CSRC_WHISPER_MEL_BANK_H_\n"
+    s += "#define KALDIFEAT_CSRC_WHISPER_MEL_BANK_H_\n"
+    s += "namespace kaldifeat {\n\n"
+    s += f"constexpr int32_t kWhisperMelRows = {m.shape[0]};\n"
+    s += f"constexpr int32_t kWhisperMelCols = {m.shape[1]};\n"
+    s += "\n"
+    s += "constexpr float kWhisperMelArray[] = {\n"
+    sep = ""
+    for i, f in enumerate(m.reshape(-1).tolist()):
+        s += f"{sep}{f:.8f}"
+        sep = ", "
+        if i and i % 7 == 0:
+            s += ",\n"
+            sep = ""
+
+    s += "};\n\n"
+    s += "}  // namespace kaldifeat\n\n"
+    s += "#endif  // KALDIFEAT_CSRC_WHISPER_MEL_BANK_H_\n"
+
+    with open("whisper-mel-bank.h", "w") as f:
+        f.write(s)
+
+
+if __name__ == "__main__":
+    main()

kaldifeat/csrc/mel-computations.cc

@@ -196,6 +196,15 @@ MelBanks::MelBanks(const MelBanksOptions &opts,
   }
 }
 
+MelBanks::MelBanks(const float *weights, int32_t num_rows, int32_t num_cols,
+                   torch::Device device)
+    : debug_(false), htk_mode_(false) {
+  bins_mat_ = torch::from_blob(const_cast<float *>(weights),
+                               {num_rows, num_cols}, torch::kFloat)
+                  .t()
+                  .to(device);
+}
+
 torch::Tensor MelBanks::Compute(const torch::Tensor &spectrum) const {
   return torch::mm(spectrum, bins_mat_);
 }

kaldifeat/csrc/mel-computations.h

@@ -76,6 +76,17 @@ class MelBanks {
            const FrameExtractionOptions &frame_opts, float vtln_warp_factor,
            torch::Device device);
 
+  // Initialize with a 2-d weights matrix
+  //
+  // Note: This constructor is for Whisper. It does not initialize
+  // center_freqs_.
+  //
+  // @param weights Pointer to the start address of the matrix
+  // @param num_rows It equals to number of mel bins
+  // @param num_cols It equals to (number of fft bins)/2+1
+  MelBanks(const float *weights, int32_t num_rows, int32_t num_cols,
+           torch::Device device);
+
   // CAUTION: we save a transposed version of bins_mat_, so return size(1) here
   int32_t NumBins() const { return static_cast<int32_t>(bins_mat_.size(1)); }
 
@@ -89,7 +100,8 @@ class MelBanks {
 
  private:
   // A 2-D matrix. Its shape is NOT [num_bins, num_fft_bins]
-  // Its shape is [num_fft_bins, num_bins].
+  // Its shape is [num_fft_bins, num_bins] for non-whisper.
+  // For whisper, its shape is [num_fft_bins/2+1, num_bins]
   torch::Tensor bins_mat_;
 
   // center frequencies of bins, numbered from 0 ... num_bins-1.

kaldifeat/csrc/whisper-fbank.cc

@@ -0,0 +1,78 @@
+/**
+ * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kaldifeat/csrc/whisper-fbank.h"
+
+#include <cmath>
+#include <vector>
+
+#include "kaldifeat/csrc/mel-computations.h"
+#include "kaldifeat/csrc/whisper-mel-bank.h"
+
+#ifndef M_2PI
+#define M_2PI 6.283185307179586476925286766559005
+#endif
+
+namespace kaldifeat {
+
+WhisperFbankComputer::WhisperFbankComputer(const WhisperFbankOptions &opts)
+    : opts_(opts),
+      mel_banks_(kWhisperMelArray, kWhisperMelRows, kWhisperMelCols,
+                 opts.device) {
+  opts_.frame_opts.samp_freq = 16000;
+  opts_.frame_opts.frame_shift_ms = 10;
+  opts_.frame_opts.frame_length_ms = 25;
+  opts_.frame_opts.dither = 0;
+  opts_.frame_opts.preemph_coeff = 0;
+  opts_.frame_opts.remove_dc_offset = false;
+  opts_.frame_opts.window_type = "hann";
+  opts_.frame_opts.round_to_power_of_two = false;
+  opts_.frame_opts.snip_edges = false;
+}
+
+torch::Tensor WhisperFbankComputer::Compute(
+    torch::Tensor /*signal_raw_log_energy*/, float /*vtln_warp*/,
+    const torch::Tensor &signal_frame) {
+  KALDIFEAT_ASSERT(signal_frame.dim() == 2);
+  KALDIFEAT_ASSERT(signal_frame.size(1) == opts_.frame_opts.PaddedWindowSize());
+
+  // note spectrum is in magnitude, not power, because of `abs()`
+#if defined(KALDIFEAT_HAS_FFT_NAMESPACE)
+  // signal_frame shape: [x, 512]
+  // power shape [x, 257]
+  torch::Tensor power = torch::fft::rfft(signal_frame).abs().pow(2);
+#else
+  // signal_frame shape [x, 512]
+  // real_imag shape [x, 257, 2],
+  //   where [..., 0] is the real part
+  //         [..., 1] is the imaginary part
+  torch::Tensor real_imag = torch::rfft(signal_frame, 1);
+  torch::Tensor real = real_imag.index({"...", 0});
+  torch::Tensor imag = real_imag.index({"...", 1});
+  torch::Tensor power = (real.square() + imag.square());
+#endif
+
+  torch::Tensor mel_energies = mel_banks_.Compute(power);
+  torch::Tensor log_spec = torch::clamp_min(mel_energies, 1e-10).log10();
+  log_spec = torch::maximum(log_spec, log_spec.max() - 8.0);
+  torch::Tensor mel = (log_spec + 4.0) / 4.0;
+
+  return mel;
+}
+
+}  // namespace kaldifeat

kaldifeat/csrc/whisper-fbank.h

@@ -0,0 +1,74 @@
+/**
+ * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDIFEAT_CSRC_WHISPER_FBANK_H_
+#define KALDIFEAT_CSRC_WHISPER_FBANK_H_
+
+#include <string>
+#include <vector>
+
+#include "kaldifeat/csrc/feature-common.h"
+#include "kaldifeat/csrc/feature-window.h"
+#include "kaldifeat/csrc/mel-computations.h"
+
+namespace kaldifeat {
+
+struct WhisperFbankOptions {
+  FrameExtractionOptions frame_opts;
+
+  torch::Device device{"cpu"};
+  std::string ToString() const {
+    std::ostringstream os;
+    os << "WhisperFbankOptions(";
+    os << "frame_opts=" << frame_opts.ToString() << ", ";
+    os << "device=\"" << device << "\")";
+    return os.str();
+  }
+};
+
+class WhisperFbankComputer {
+ public:
+  // note: Only frame_opts.device is used. All other fields from frame_opts
+  // are ignored
+  explicit WhisperFbankComputer(const WhisperFbankOptions &opts = {});
+
+  int32_t Dim() const { return 80; }
+
+  const FrameExtractionOptions &GetFrameOptions() const {
+    return opts_.frame_opts;
+  }
+
+  const WhisperFbankOptions &GetOptions() const { return opts_; }
+
+  torch::Tensor Compute(torch::Tensor /*signal_raw_log_energy*/,
+                        float /*vtln_warp*/, const torch::Tensor &signal_frame);
+
+  // if true, compute log_energy_pre_window but after dithering and dc removal
+  bool NeedRawLogEnergy() const { return false; }
+  using Options = WhisperFbankOptions;
+
+ private:
+  WhisperFbankOptions opts_;
+  MelBanks mel_banks_;
+};
+
+using WhisperFbank = OfflineFeatureTpl<WhisperFbankComputer>;
+
+}  // namespace kaldifeat
+
+#endif  // KALDIFEAT_CSRC_WHISPER_FBANK_H_