[Op] ISTFT op class, shape infer and reference

src/core/include/openvino/op/istft.hpp

@@ -0,0 +1,67 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/op/op.hpp"
+
+namespace ov {
+namespace op {
+namespace v16 {
+/// \brief An operation ISTFT that computes the Inverse Short Time Fourier Transform.
+/// \ingroup ov_ops_cpp_api
+class OPENVINO_API ISTFT : public Op {
+public:
+    OPENVINO_OP("ISTFT", "opset16");
+    ISTFT() = default;
+
+    /// \brief Constructs an ISTFT operation with signal length to be inferred
+    ///
+    /// \param data  Input data
+    /// \param window Window values applied in ISTFT
+    /// \param frame_size Scalar value representing the size of Fourier Transform
+    /// \param frame_step The distance (number of samples) between successive window frames
+    /// \param center Flag signaling if the signal input has been padded before STFT
+    /// \param normalized Flag signaling if the STFT result has been normalized
+    ISTFT(const Output<Node>& data,
+          const Output<Node>& window,
+          const Output<Node>& frame_size,
+          const Output<Node>& frame_step,
+          const bool center,
+          const bool normalized);
+
+    /// \brief Constructs an ISTFT operation with signal length provided
+    ///
+    /// \param data  Input data
+    /// \param window Window values applied in ISTFT
+    /// \param frame_size Scalar value representing the size of Fourier Transform
+    /// \param frame_step The distance (number of samples) between successive window frames
+    /// \param signal_length The signal length of the original signal
+    /// \param center Flag signaling if the signal input has been padded before STFT
+    /// \param normalized Flag signaling if the STFT result has been normalized
+    ISTFT(const Output<Node>& data,
+          const Output<Node>& window,
+          const Output<Node>& frame_size,
+          const Output<Node>& frame_step,
+          const Output<Node>& signal_length,
+          const bool center,
+          const bool normalized);
+
+    bool visit_attributes(AttributeVisitor& visitor) override;
+    void validate_and_infer_types() override;
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+
+    bool get_center() const;
+    void set_center(const bool center);
+
+    bool get_normalized() const;
+    void set_normalized(const bool normalized);
+
+private:
+    bool m_center = false;
+    bool m_normalized = false;
+};
+}  // namespace v16
+}  // namespace op
+}  // namespace ov

src/core/include/openvino/op/ops.hpp

@@ -99,6 +99,7 @@
 #include "openvino/op/is_finite.hpp"
 #include "openvino/op/is_inf.hpp"
 #include "openvino/op/is_nan.hpp"
+#include "openvino/op/istft.hpp"
 #include "openvino/op/less.hpp"
 #include "openvino/op/less_eq.hpp"
 #include "openvino/op/log.hpp"

src/core/include/openvino/opsets/opset16_tbl.hpp

@@ -15,3 +15,4 @@ _OPENVINO_OP_REG(ShapeOf, ov::op::v3)
 
 // New operations added in opset16
 _OPENVINO_OP_REG(Identity, ov::op::v16)
+_OPENVINO_OP_REG(ISTFT, ov::op::v16)

src/core/reference/include/openvino/reference/istft.hpp

@@ -0,0 +1,22 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/core/shape.hpp"
+
+namespace ov {
+namespace reference {
+void istft(const float* in_data,
+           const float* window,
+           float* final_result,
+           const Shape& signal_shape,
+           const Shape& window_shape,
+           const int64_t frame_size,
+           const int64_t frame_step,
+           const int64_t length,
+           const bool center,
+           const bool normalized);
+}  // namespace reference
+}  // namespace ov

src/core/reference/src/op/istft.cpp

@@ -0,0 +1,135 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "openvino/reference/istft.hpp"
+
+#include <algorithm>
+#include <functional>
+#include <vector>
+
+#include "openvino/core/shape.hpp"
+#include "openvino/reference/irdft.hpp"
+#include "openvino/reference/transpose.hpp"
+
+namespace ov {
+namespace reference {
+void istft(const float* in_data,
+           const float* window,
+           float* final_result,
+           const Shape& data_shape,
+           const Shape& window_shape,
+           const int64_t frame_size,
+           const int64_t frame_step,
+           const int64_t length,
+           const bool center,
+           const bool normalized) {
+    const auto is_data_3D = data_shape.size() == 3;
+    const size_t frames_axis = 1 + (is_data_3D ? 0 : 1);
+    const size_t batch_size = is_data_3D ? 1 : data_shape[0];
+
+    const auto sqrt_frame_size = static_cast<float>(std::sqrt(frame_size));
+    const auto num_frames = data_shape[frames_axis];
+
+    const auto signal_length = (num_frames - 1) * frame_step + frame_size;
+    const int64_t final_signal_length = length > 0 ? length : (center ? (signal_length - frame_size) : signal_length);
+
+    std::vector<float> mid_result(batch_size * signal_length, 0);
+    float* result = mid_result.data();
+
+    const auto frame_size_dim = static_cast<size_t>(frame_size);
+    const auto frame_size_dim_shape = Shape{frame_size_dim};
+    const auto frame_size_dim_shape_out = Shape{frame_size_dim, 2};
+    const auto fft_out_shape = Shape{static_cast<size_t>((frame_size_dim / 2) + 1), 2};
+
+    const auto window_length = window_shape[0] < frame_size_dim ? window_shape[0] : frame_size_dim;
+    std::vector<float> pad_window(frame_size, 0);
+    std::copy(window, window + window_shape[0], pad_window.begin() + (frame_size_dim - window_length) / 2);
+
+    const bool transpose_frames = true;
+    std::vector<float> data_t(in_data, in_data + shape_size(data_shape));
+    if (transpose_frames) {
+        const auto stft_transp_out_shape = Shape{batch_size, num_frames, fft_out_shape[0], fft_out_shape[1]};
+        transpose(reinterpret_cast<const char*>(in_data),
+                  reinterpret_cast<char*>(data_t.data()),
+                  Shape{batch_size, fft_out_shape[0], num_frames, fft_out_shape[1]},
+                  sizeof(float),
+                  {0, 2, 1, 3},
+                  stft_transp_out_shape);
+    }
+
+    const auto fft_out_shape_size = shape_size(fft_out_shape);
+    std::vector<float> window_sum(batch_size * signal_length);
+
+    for (size_t batch = 0, batch_in_start = 0, batch_out_start = 0; batch < batch_size; ++batch) {
+        for (size_t frame_idx = 0; frame_idx < num_frames; ++frame_idx) {
+            const auto in_frame_start = batch_in_start + frame_idx * fft_out_shape_size;
+            const auto in_frame_end = in_frame_start + fft_out_shape_size;
+
+            const auto out_frame_start = batch_out_start + frame_idx * frame_step;
+            const auto out_frame_end = out_frame_start + frame_size;
+
+            std::vector<float> frame_data(data_t.data() + in_frame_start, data_t.data() + in_frame_end);
+            std::vector<float> frame_signal(frame_size);
+
+            reference::irdft(frame_data,
+                             fft_out_shape,
+                             {0},
+                             frame_signal.data(),
+                             frame_size_dim_shape_out,
+                             frame_size_dim_shape,
+                             frame_size);
+
+            std::transform(frame_signal.begin(),
+                           frame_signal.end(),
+                           mid_result.begin() + out_frame_start,
+                           mid_result.begin() + out_frame_start,
+                           std::plus<float>());
+
+            std::transform(window_sum.begin() + out_frame_start,
+                           window_sum.begin() + out_frame_end,
+                           pad_window.begin(),
+                           window_sum.begin() + out_frame_start,
+                           std::plus<float>());
+        }
+
+        if (normalized) {
+            std::transform(result + batch_out_start,
+                           result + batch_out_start + signal_length,
+                           result + batch_out_start,
+                           [sqrt_frame_size](float a) {
+                               return a * sqrt_frame_size;
+                           });
+        }
+
+        std::transform(result + batch_out_start,
+                       result + batch_out_start + signal_length,
+                       window_sum.begin(),
+                       result + batch_out_start,
+                       [](float a, float b) {
+                           if (b != 0.f)
+                               return a / b;
+                           else
+                               return 0.f;
+                       });
+
+        if (center) {
+            const int64_t margin = (frame_size / 2);
+            const size_t result_start = batch_out_start + margin;
+            const int64_t data_end = signal_length - (frame_size / 2);
+            int64_t signal_end = final_signal_length < data_end ? final_signal_length : data_end;
+            std::copy(result + result_start,
+                      result + result_start + signal_end,
+                      final_result + (batch * final_signal_length));
+        } else {
+            std::copy(result + batch_out_start,
+                      result + batch_out_start + final_signal_length,
+                      final_result + batch_out_start);
+        }
+
+        batch_in_start += (num_frames * fft_out_shape_size);
+        batch_out_start += signal_length;
+    }
+}
+}  // namespace reference
+}  // namespace ov

src/core/shape_inference/include/istft_shape_inference.hpp

@@ -0,0 +1,128 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "dimension_util.hpp"
+#include "openvino/op/istft.hpp"
+#include "utils.hpp"
+
+namespace ov {
+namespace op {
+namespace v16 {
+template <class TShape, class TRShape = result_shape_t<TShape>>
+std::vector<TRShape> shape_infer(const ISTFT* op,
+                                 const std::vector<TShape>& input_shapes,
+                                 const ITensorAccessor& ta = make_tensor_accessor()) {
+    using TDim = typename TRShape::value_type;
+    using TDimVal = typename TDim::value_type;
+
+    const auto inputs_count = input_shapes.size();
+    const auto is_in_count_correct = inputs_count == 4 || inputs_count == 5;
+    NODE_VALIDATION_CHECK(op, is_in_count_correct);
+
+    const auto& data_shape = input_shapes[0];
+    const auto& window_shape = input_shapes[1];
+    const auto& frame_size_shape = input_shapes[2];
+    const auto& frame_step_shape = input_shapes[3];
+
+    const auto data_shape_rank = data_shape.rank();
+    NODE_SHAPE_INFER_CHECK(op,
+                           input_shapes,
+                           data_shape_rank.compatible(3) || data_shape_rank.compatible(4),
+                           "The shape of data must be 3D or 4D.");
+    NODE_SHAPE_INFER_CHECK(op,
+                           input_shapes,
+                           window_shape.rank().compatible(1),
+                           "The shape of window must be 1D [window_size].");
+    NODE_SHAPE_INFER_CHECK(op,
+                           input_shapes,
+                           frame_size_shape.rank().compatible(0),
+                           "The shape of frame_size must be a scalar.");
+    NODE_SHAPE_INFER_CHECK(op,
+                           input_shapes,
+                           frame_step_shape.rank().compatible(0),
+                           "The shape of frame_step must be a scalar.");
+
+    const auto frame_size = get_input_const_data_as<TRShape, int64_t>(op, 2, ta);
+    const auto frame_step = get_input_const_data_as<TRShape, int64_t>(op, 3, ta);
+
+    if (frame_size) {
+        const auto& frame_size_val = (*frame_size)[0];
+        NODE_SHAPE_INFER_CHECK(op,
+                               input_shapes,
+                               0 < frame_size_val,
+                               "Provided frame size is ",
+                               frame_size_val,
+                               " but must be greater than zero.");
+        const bool is_win_shape_correct =
+            window_shape.is_dynamic() || (TDimVal{0} < window_shape[0].get_length() &&
+                                          window_shape[0].get_length() <= static_cast<TDimVal>(frame_size_val));
+
+        NODE_SHAPE_INFER_CHECK(op,
+                               input_shapes,
+                               is_win_shape_correct,
+                               "Window input dimension must be in range [1, ",
+                               frame_size_val,
+                               "].");
+    }
+
+    if (frame_step) {
+        const auto& frame_step_val = (*frame_step)[0];
+        NODE_SHAPE_INFER_CHECK(op,
+                               input_shapes,
+                               0 < frame_step_val,
+                               "Provided frame step is ",
+                               frame_step_val,
+                               " but must be greater than zero.");
+    }
+
+    // For the input with dynamic rank, output shape is also fully dynamic
+    if (data_shape_rank.is_dynamic()) {
+        return {data_shape};
+    }
+    const auto is_data_3D = data_shape.size() == 3;
+
+    std::vector<TRShape> output_shapes;
+    if (inputs_count == 5) {
+        const auto& length_shape = input_shapes[4];
+        const bool has_len_valid_shape =
+            length_shape.rank().is_dynamic() ||
+            (length_shape.size() == 0 || (length_shape.size() == 1 && length_shape[0].compatible(1)));
+        NODE_SHAPE_INFER_CHECK(op,
+                               input_shapes,
+                               has_len_valid_shape,
+                               "The shape of 'signal_length' input must be a scalar or single element 1D tensor.");
+
+        const auto sig_len_in = get_input_const_data_as_shape<TRShape>(op, 4, ta);
+        if (sig_len_in) {  // Set desired length of the signal dimension, if provided
+            output_shapes.emplace_back(TRShape{(*sig_len_in)[0]});
+        } else {
+            output_shapes.emplace_back(TRShape{TDim(ov::util::dim::inf_bound)});
+        }
+    } else if (frame_size && frame_step) {  // Otherwise infer the length of the signal
+        const auto& frame_size_val = (*frame_size)[0];
+        const auto& frame_step_val = (*frame_step)[0];
+
+        const int64_t frames_axis = 1 + (is_data_3D ? 0 : 1);
+        const TDim& num_frames_dim = data_shape[frames_axis];
+        TDim signal_length = (num_frames_dim - 1) * frame_step_val;
+        if (!op->get_center()) {
+            signal_length += frame_size_val;
+        }
+        output_shapes.emplace_back(TRShape{std::move(signal_length)});
+    } else {  // Not enough info to infer the signal lenght, set dynamic dimension
+        output_shapes.emplace_back(TRShape{TDim(ov::util::dim::inf_bound)});
+    }
+
+    if (!is_data_3D) {  // Copy batch dimension
+        const auto& batch_dim = data_shape[0];
+        output_shapes[0].insert(output_shapes[0].begin(), batch_dim);
+    }
+
+    return output_shapes;
+}
+}  // namespace v16
+}  // namespace op
+}  // namespace ov