Add transducer modified_beam_search for RKNN. (#1949)

sherpa-onnx/csrc/CMakeLists.txt

@@ -155,6 +155,7 @@ if(SHERPA_ONNX_ENABLE_RKNN)
   list(APPEND sources
     ./rknn/online-stream-rknn.cc
     ./rknn/online-transducer-greedy-search-decoder-rknn.cc
+    ./rknn/online-transducer-modified-beam-search-decoder-rknn.cc
     ./rknn/online-zipformer-ctc-model-rknn.cc
     ./rknn/online-zipformer-transducer-model-rknn.cc
     ./rknn/utils.cc

sherpa-onnx/csrc/hypothesis.h

@@ -142,7 +142,6 @@ class Hypotheses {
 
   void Clear() { hyps_dict_.clear(); }
 
- private:
   // Return a list of hyps contained in this object.
   std::vector<Hypothesis> Vec() const {
     std::vector<Hypothesis> ans;

sherpa-onnx/csrc/math.h

@@ -119,5 +119,17 @@ std::vector<int32_t> TopkIndex(const T *vec, int32_t size, int32_t topk) {
   return {vec_index.begin(), vec_index.begin() + k_num};
 }
 
+template <class T>
+std::vector<int32_t> TopkIndex(const std::vector<std::vector<T>> &vec,
+                               int32_t topk) {
+  std::vector<T> flatten;
+  flatten.reserve(vec.size() * vec[0].size());
+  for (const auto &v : vec) {
+    flatten.insert(flatten.end(), v.begin(), v.end());
+  }
+
+  return TopkIndex(flatten.data(), flatten.size(), topk);
+}
+
 }  // namespace sherpa_onnx
 #endif  // SHERPA_ONNX_CSRC_MATH_H_

sherpa-onnx/csrc/rknn/online-recognizer-transducer-rknn-impl.h

@@ -16,7 +16,9 @@
 #include "sherpa-onnx/csrc/online-recognizer-impl.h"
 #include "sherpa-onnx/csrc/online-recognizer.h"
 #include "sherpa-onnx/csrc/rknn/online-stream-rknn.h"
+#include "sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h"
 #include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
+#include "sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.h"
 #include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
 #include "sherpa-onnx/csrc/symbol-table.h"
 
@@ -87,8 +89,20 @@ class OnlineRecognizerTransducerRknnImpl : public OnlineRecognizerImpl {
       unk_id_ = sym_["<unk>"];
     }
 
+    if (config.decoding_method == "greedy_search") {
       decoder_ = std::make_unique<OnlineTransducerGreedySearchDecoderRknn>(
           model_.get(), unk_id_);
+    } else if (config.decoding_method == "modified_beam_search") {
+      decoder_ =
+          std::make_unique<OnlineTransducerModifiedBeamSearchDecoderRknn>(
+              model_.get(), config.max_active_paths, unk_id_);
+    } else {
+      SHERPA_ONNX_LOGE(
+          "Invalid decoding method: '%s'. Support only greedy_search and "
+          "modified_beam_search.",
+          config.decoding_method.c_str());
+      SHERPA_ONNX_EXIT(-1);
+    }
   }
 
   template <typename Manager>
@@ -223,7 +237,7 @@ class OnlineRecognizerTransducerRknnImpl : public OnlineRecognizerImpl {
   Endpoint endpoint_;
   int32_t unk_id_ = -1;
   std::unique_ptr<OnlineZipformerTransducerModelRknn> model_;
-  std::unique_ptr<OnlineTransducerGreedySearchDecoderRknn> decoder_;
+  std::unique_ptr<OnlineTransducerDecoderRknn> decoder_;
 };
 
 }  // namespace sherpa_onnx

sherpa-onnx/csrc/rknn/online-stream-rknn.h

@@ -8,7 +8,7 @@
 
 #include "rknn_api.h"  // NOLINT
 #include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
+#include "sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h"
 
 namespace sherpa_onnx {
 

sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h

@@ -0,0 +1,63 @@
+// sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_
+#define SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_
+
+#include <vector>
+
+#include "sherpa-onnx/csrc/hypothesis.h"
+#include "sherpa-onnx/csrc/macros.h"
+
+namespace sherpa_onnx {
+
+struct OnlineTransducerDecoderResultRknn {
+  /// Number of frames after subsampling we have decoded so far
+  int32_t frame_offset = 0;
+
+  /// The decoded token IDs so far
+  std::vector<int64_t> tokens;
+
+  /// number of trailing blank frames decoded so far
+  int32_t num_trailing_blanks = 0;
+
+  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
+  std::vector<int32_t> timestamps;
+
+  // used only by greedy_search
+  std::vector<float> previous_decoder_out;
+
+  // used only in modified beam_search
+  Hypotheses hyps;
+
+  // used only by modified_beam_search
+  std::vector<std::vector<float>> previous_decoder_out2;
+};
+
+class OnlineTransducerDecoderRknn {
+ public:
+  virtual ~OnlineTransducerDecoderRknn() = default;
+
+  /* Return an empty result.
+   *
+   * To simplify the decoding code, we add `context_size` blanks
+   * to the beginning of the decoding result, which will be
+   * stripped by calling `StripPrecedingBlanks()`.
+   */
+  virtual OnlineTransducerDecoderResultRknn GetEmptyResult() const = 0;
+
+  /** Strip blanks added by `GetEmptyResult()`.
+   *
+   * @param r It is changed in-place.
+   */
+  virtual void StripLeadingBlanks(
+      OnlineTransducerDecoderResultRknn * /*r*/) const {}
+
+  virtual void Decode(std::vector<float> encoder_out,
+                      OnlineTransducerDecoderResultRknn *result) const = 0;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_

sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h

@@ -7,39 +7,26 @@
 
 #include <vector>
 
+#include "sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h"
+#include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
 #include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
 
 namespace sherpa_onnx {
 
-struct OnlineTransducerDecoderResultRknn {
+class OnlineTransducerGreedySearchDecoderRknn
-  /// Number of frames after subsampling we have decoded so far
+    : public OnlineTransducerDecoderRknn {
-  int32_t frame_offset = 0;
-
-  /// The decoded token IDs so far
-  std::vector<int64_t> tokens;
-
-  /// number of trailing blank frames decoded so far
-  int32_t num_trailing_blanks = 0;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  std::vector<int32_t> timestamps;
-
-  std::vector<float> previous_decoder_out;
-};
-
-class OnlineTransducerGreedySearchDecoderRknn {
  public:
   explicit OnlineTransducerGreedySearchDecoderRknn(
       OnlineZipformerTransducerModelRknn *model, int32_t unk_id = 2,
       float blank_penalty = 0.0)
       : model_(model), unk_id_(unk_id), blank_penalty_(blank_penalty) {}
 
-  OnlineTransducerDecoderResultRknn GetEmptyResult() const;
+  OnlineTransducerDecoderResultRknn GetEmptyResult() const override;
 
-  void StripLeadingBlanks(OnlineTransducerDecoderResultRknn *r) const;
+  void StripLeadingBlanks(OnlineTransducerDecoderResultRknn *r) const override;
 
   void Decode(std::vector<float> encoder_out,
-              OnlineTransducerDecoderResultRknn *result) const;
+              OnlineTransducerDecoderResultRknn *result) const override;
 
  private:
   OnlineZipformerTransducerModelRknn *model_;  // Not owned

sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.cc

@@ -0,0 +1,146 @@
+// sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.cc
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.h"
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include "sherpa-onnx/csrc/hypothesis.h"
+#include "sherpa-onnx/csrc/macros.h"
+#include "sherpa-onnx/csrc/math.h"
+
+namespace sherpa_onnx {
+
+OnlineTransducerDecoderResultRknn
+OnlineTransducerModifiedBeamSearchDecoderRknn::GetEmptyResult() const {
+  int32_t context_size = model_->ContextSize();
+  int32_t blank_id = 0;  // always 0
+  OnlineTransducerDecoderResultRknn r;
+
+  std::vector<int64_t> blanks(context_size, -1);
+  blanks.back() = blank_id;
+
+  Hypotheses blank_hyp({{blanks, 0}});
+  r.hyps = std::move(blank_hyp);
+  r.tokens = std::move(blanks);
+
+  return r;
+}
+
+void OnlineTransducerModifiedBeamSearchDecoderRknn::StripLeadingBlanks(
+    OnlineTransducerDecoderResultRknn *r) const {
+  int32_t context_size = model_->ContextSize();
+  auto hyp = r->hyps.GetMostProbable(true);
+
+  std::vector<int64_t> tokens(hyp.ys.begin() + context_size, hyp.ys.end());
+  r->tokens = std::move(tokens);
+  r->timestamps = std::move(hyp.timestamps);
+
+  r->num_trailing_blanks = hyp.num_trailing_blanks;
+}
+
+static std::vector<std::vector<float>> GetDecoderOut(
+    OnlineZipformerTransducerModelRknn *model, const Hypotheses &hyp_vec) {
+  std::vector<std::vector<float>> ans;
+  ans.reserve(hyp_vec.Size());
+
+  int32_t context_size = model->ContextSize();
+  for (const auto &p : hyp_vec) {
+    const auto &hyp = p.second;
+    auto start = hyp.ys.begin() + (hyp.ys.size() - context_size);
+    auto end = hyp.ys.end();
+    auto tokens = std::vector<int64_t>(start, end);
+    auto decoder_out = model->RunDecoder(std::move(tokens));
+
+    ans.push_back(std::move(decoder_out));
+  }
+
+  return ans;
+}
+
+static std::vector<std::vector<float>> GetJoinerOutLogSoftmax(
+    OnlineZipformerTransducerModelRknn *model, const float *p_encoder_out,
+    const std::vector<std::vector<float>> &decoder_out) {
+  std::vector<std::vector<float>> ans;
+  ans.reserve(decoder_out.size());
+
+  for (const auto &d : decoder_out) {
+    auto joiner_out = model->RunJoiner(p_encoder_out, d.data());
+
+    LogSoftmax(joiner_out.data(), joiner_out.size());
+
+    ans.push_back(std::move(joiner_out));
+  }
+  return ans;
+}
+
+void OnlineTransducerModifiedBeamSearchDecoderRknn::Decode(
+    std::vector<float> encoder_out,
+    OnlineTransducerDecoderResultRknn *result) const {
+  auto &r = result[0];
+  auto attr = model_->GetEncoderOutAttr();
+  int32_t num_frames = attr.dims[1];
+  int32_t encoder_out_dim = attr.dims[2];
+
+  int32_t vocab_size = model_->VocabSize();
+  int32_t context_size = model_->ContextSize();
+
+  Hypotheses cur = std::move(result->hyps);
+  std::vector<Hypothesis> prev;
+
+  auto decoder_out = std::move(result->previous_decoder_out2);
+  if (decoder_out.empty()) {
+    decoder_out = GetDecoderOut(model_, cur);
+  }
+
+  const float *p_encoder_out = encoder_out.data();
+
+  int32_t frame_offset = result->frame_offset;
+
+  for (int32_t t = 0; t != num_frames; ++t) {
+    prev = cur.Vec();
+    cur.Clear();
+
+    auto log_probs = GetJoinerOutLogSoftmax(model_, p_encoder_out, decoder_out);
+    p_encoder_out += encoder_out_dim;
+
+    for (int32_t i = 0; i != prev.size(); ++i) {
+      auto log_prob = prev[i].log_prob;
+      for (auto &p : log_probs[i]) {
+        p += log_prob;
+      }
+    }
+
+    auto topk = TopkIndex(log_probs, max_active_paths_);
+    for (auto k : topk) {
+      int32_t hyp_index = k / vocab_size;
+      int32_t new_token = k % vocab_size;
+
+      Hypothesis new_hyp = prev[hyp_index];
+      new_hyp.log_prob = log_probs[hyp_index][new_token];
+
+      // blank is hardcoded to 0
+      // also, it treats unk as blank
+      if (new_token != 0 && new_token != unk_id_) {
+        new_hyp.ys.push_back(new_token);
+        new_hyp.timestamps.push_back(t + frame_offset);
+        new_hyp.num_trailing_blanks = 0;
+
+      } else {
+        ++new_hyp.num_trailing_blanks;
+      }
+      cur.Add(std::move(new_hyp));
+    }
+
+    decoder_out = GetDecoderOut(model_, cur);
+  }
+
+  result->hyps = std::move(cur);
+  result->frame_offset += num_frames;
+  result->previous_decoder_out2 = std::move(decoder_out);
+}
+
+}  // namespace sherpa_onnx

sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.h

@@ -0,0 +1,42 @@
+// sherpa-onnx/csrc/rknn/online-transducer-modified-beam-search-decoder-rknn.h
+//
+// Copyright (c)  2025  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_RKNN_H_
+#define SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_RKNN_H_
+
+#include <vector>
+
+#include "sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h"
+#include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
+
+namespace sherpa_onnx {
+
+class OnlineTransducerModifiedBeamSearchDecoderRknn
+    : public OnlineTransducerDecoderRknn {
+ public:
+  explicit OnlineTransducerModifiedBeamSearchDecoderRknn(
+      OnlineZipformerTransducerModelRknn *model, int32_t max_active_paths,
+      int32_t unk_id = 2, float blank_penalty = 0.0)
+      : model_(model),
+        max_active_paths_(max_active_paths),
+        unk_id_(unk_id),
+        blank_penalty_(blank_penalty) {}
+
+  OnlineTransducerDecoderResultRknn GetEmptyResult() const override;
+
+  void StripLeadingBlanks(OnlineTransducerDecoderResultRknn *r) const override;
+
+  void Decode(std::vector<float> encoder_out,
+              OnlineTransducerDecoderResultRknn *result) const override;
+
+ private:
+  OnlineZipformerTransducerModelRknn *model_;  // Not owned
+  int32_t max_active_paths_;
+  int32_t unk_id_;
+  float blank_penalty_;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_RKNN_H_

sherpa-onnx/csrc/rknn/utils.cc

@@ -6,6 +6,7 @@
 
 #include <sstream>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "sherpa-onnx/csrc/macros.h"