Upload the AX630C inference workflow.

.gitattributes

@@ -39,6 +39,7 @@ cpp/install/ax650/zipvoice filter=lfs diff=lfs merge=lfs -text
 cpp/install/ax650/zipvoice_axera filter=lfs diff=lfs merge=lfs -text
 cpp/install/axcl/zipvoice filter=lfs diff=lfs merge=lfs -text
 cpp/vocoder/vocos_full.axmodel filter=lfs diff=lfs merge=lfs -text
+cpp/vocoder/vocos_full_ax630c.axmodel filter=lfs diff=lfs merge=lfs -text
 models/axmodels/decoder_part0.axmodel filter=lfs diff=lfs merge=lfs -text
 models/axmodels/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
 models/axmodels/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
@@ -53,6 +54,10 @@ models/zipvoice_distill_ax630C/decoder_part0.axmodel filter=lfs diff=lfs merge=l
 models/zipvoice_distill_ax630C/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/decoder_part3.axmodel filter=lfs diff=lfs merge=lfs -text
+models/zipvoice_distill_ax630C/decoder_part3_0.axmodel filter=lfs diff=lfs merge=lfs -text
+models/zipvoice_distill_ax630C/decoder_part3_1.axmodel filter=lfs diff=lfs merge=lfs -text
+models/zipvoice_distill_ax630C/decoder_part3_2.axmodel filter=lfs diff=lfs merge=lfs -text
+models/zipvoice_distill_ax630C/decoder_part3_3.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/encoder.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part0.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text

README.md

@@ -18,12 +18,12 @@ ZipVoice Distill 是 ZipVoice 的蒸馏版本，主要优势是在较小性能
 
 ### Distill 模型 (num_step=4)
 
-| 场景 | 音频时长 | Python AX650 (耗时/RTF) | C++ AX650 (耗时/RTF) | C++ AXCL (耗时/RTF) |
-|------|---------|------------------------|---------------------|---------------------|
-| 中文句子 | 6.41s | 1.992s / 0.311 | 1.057s / 0.165 | 0.918s / **0.143** |
-| 中文段落 | 44.97s | 13.457s / 0.301 | 7.357s / 0.164 | 6.443s / **0.143** |
-| 英文句子 | 6.41s | 2.045s / 0.319 | 1.067s / 0.166 | 0.919s / **0.143** |
-| 英文段落 | 59.16s | 19.715s / 0.305 | 10.529s / 0.178 | 9.207s / **0.156** |
+| 场景 | 音频时长 | Python AX650<br>耗时 RTF | C++ AX650<br>耗时 RTF | C++ AXCL<br>耗时 RTF | Python AX630C<br>耗时 RTF | C++ AX630C<br>耗时 RTF |
+|------|---------|-------------------------|----------------------|---------------------|--------------------------|-----------------------|
+| 中文句子 | 6.41s | 1.992s/0.311 | 1.057s/0.165 | 0.918s/**0.143** | 10.296s/1.606 | 4.650s/0.725 |
+| 中文段落 | 44.97s | 13.457s/0.301 | 7.357s/0.164 | 6.443s/**0.143** | 71.574s/1.600 | 32.590s/0.725 |
+| 英文句子 | 6.41s | 2.045s/0.319 | 1.067s/0.166 | 0.919s/**0.143** | 10.686s/1.667 | 4.654s/0.726 |
+| 英文段落 | 59.16s | 19.715s/0.305 | 10.529s/0.178 | 9.207s/**0.156** | 106.183s/1.640 | 46.541s/0.787 |
 
 ### 普通模型 (num_step=10)
 
@@ -53,7 +53,7 @@ ZipVoice Distill 是 ZipVoice 的蒸馏版本，主要优势是在较小性能
 ## 目录结构
 
 ```text
-ZipVoice.AXERA_AXCL/
+ZipVoice.AXERA/
 ├── assets/
 │   ├── moss_prompts/
 │   └── paragraphs/
@@ -67,8 +67,9 @@ ZipVoice.AXERA_AXCL/
 ├── cpp/                         # C++ 源码 (AXERA & AXCL)
 ├── scripts/
 ├── infer_zipvoice_axera.py      # Python 推理入口
-├── run_ax650.sh                 # AXERA 板端一键运行
-├── run_axcl.sh                  # AXCL 算力卡一键运行
+├── run_ax650.sh                 # AX650 板端
+├── run_ax630c.sh                # AX630C 板端
+├── run_axcl.sh                  # AXCL 算力卡
 ├── requirements.txt
 └── README.md
 ```
@@ -279,10 +280,21 @@ RTF: 0.3045
 音频：[outputs/en_long_paragraph_distill_ax650.wav](outputs/en_long_paragraph_distill_ax650.wav)  
 提示音：[assets/moss_prompts/en_4_4p5s.wav](assets/moss_prompts/en_4_4p5s.wav)
 
+#### AX630C ZipVoice Distill
+
+普通模型推理太慢，这边仅提供 Distill 模型，参数与 AX650 一致，仅 `--model-name` 改为 `zipvoice_distill_ax630C`：
+
+```bash
+python3 infer_zipvoice_axera.py \
+  --model-name zipvoice_distill_ax630C \
+  --text "今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。" \
+  --prompt-text "不管怎么样我和汤姆还是要感谢贝尔卡金的援手" \
+  --prompt-wav assets/moss_prompts/zh_1_4p5s.wav --seed 42
+```
 
 ### 参数说明
 
-- `--model-name`：选择模型目录。可选 `zipvoice_ax650`、`zipvoice_distill_ax650`、`zipvoice_distill_ax630C`。
+- `--model-name`：`zipvoice_ax650` / `zipvoice_distill_ax650` / `zipvoice_distill_ax630C`。
 - `--prompt-wav`：参考音频，用于控制音色，建议 3-5s。
 - `--prompt-text`：参考音频对应文本，必须尽量和 `prompt-wav` 内容一致。
 - `--num-step`：采样步数。默认从模型目录的 `runtime_config.json` 读取。

cpp/README.md

@@ -4,12 +4,12 @@
 
 ### Distill 模型 (num_step=4)
 
-| 场景 | 音频时长 | Python AX650 (耗时/RTF) | C++ AX650 (耗时/RTF) | C++ AXCL (耗时/RTF) |
-|------|---------|------------------------|---------------------|---------------------|
-| 中文句子 | 6.41s | 1.992s / 0.311 | 1.057s / 0.165 | 0.918s / **0.143** |
-| 中文段落 | 44.97s | 13.457s / 0.301 | 7.357s / 0.164 | 6.443s / **0.143** |
-| 英文句子 | 6.41s | 2.045s / 0.319 | 1.067s / 0.166 | 0.919s / **0.143** |
-| 英文段落 | 59.16s | 19.715s / 0.305 | 10.529s / 0.178 | 9.207s / **0.156** |
+| 场景 | 音频 | Python AX650<br>耗时 RTF | C++ AX650<br>耗时 RTF | C++ AXCL<br>耗时 RTF | Python AX630C<br>耗时 RTF | C++ AX630C<br>耗时 RTF |
+|------|------|-------------------------|----------------------|---------------------|--------------------------|-----------------------|
+| 中文句子 | 6.41s | 1.992s/0.311 | 1.057s/0.165 | 0.918s/**0.143** | 10.296s/1.606 | 4.650s/0.725 |
+| 中文段落 | 44.97s | 13.457s/0.301 | 7.357s/0.164 | 6.443s/**0.143** | 71.574s/1.600 | 32.590s/0.725 |
+| 英文句子 | 6.41s | 2.045s/0.319 | 1.067s/0.166 | 0.919s/**0.143** | 10.686s/1.667 | 4.654s/0.726 |
+| 英文段落 | 59.16s | 19.715s/0.305 | 10.529s/0.178 | 9.207s/**0.156** | 106.183s/1.640 | 46.541s/0.787 |
 
 ### 普通模型 (num_step=10)
 
@@ -32,6 +32,7 @@ mkdir -p <TOOLCHAIN_DIR> && cd <TOOLCHAIN_DIR>
 wget https://developer.arm.com/-/media/Files/downloads/gnu-a/9.2-2019.12/binrel/gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu.tar.xz
 tar -xf gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu.tar.xz
 git clone https://github.com/AXERA-TECH/ax650n_bsp_sdk.git --depth=1
+git clone https://github.com/AXERA-TECH/ax620e_bsp_sdk.git --depth=1
 git clone https://github.com/Abandon-ht/axcl_bsp_sdk.git --depth=1
 ```
 
@@ -46,6 +47,7 @@ pip install jieba numpy pypinyin piper_phonemize
 
 ```bash
 bash build_ax650.sh    # → install/ax650/zipvoice
+bash build_ax630c.sh   # → install/ax630c/zipvoice
 bash build_axcl.sh     # → install/axcl/zipvoice
 ```
 
@@ -67,6 +69,10 @@ bash run_ax650.sh standard en paragraph
 # 算力卡
 bash run_axcl.sh distill zh paragraph
 bash run_axcl.sh standard en sentence
+
+# AX630C (仅 distill)
+bash run_ax630c.sh zh sentence
+bash run_ax630c.sh en paragraph
 ...
 ```
 

cpp/cmake/msp_dependencies.cmake

@@ -19,10 +19,7 @@ if (AX650)
 elseif(AX630C)
     add_definitions(-DAX630C)
     if(NOT BSP_MSP_DIR)
-        set(BSP_MSP_DIR ${CMAKE_SOURCE_DIR}/ax620e_bsp_sdk/msp/out/arm64_glibc)
-    endif()
-    if(NOT EXISTS ${BSP_MSP_DIR})
-        message(FATAL_ERROR "BSP_MSP_DIR ${BSP_MSP_DIR} not exist.")
+        set(BSP_MSP_DIR /data/shared/huyuan/toolchains/ax620e_bsp_sdk/msp/out/arm64_glibc)
     endif()
     set(MSP_INC_DIR ${BSP_MSP_DIR}/include)
     set(MSP_LIB_DIR ${BSP_MSP_DIR}/lib)

cpp/download_bsp.sh

@@ -1,6 +1,5 @@
 #!/bin/bash
 # Download BSP SDKs to /data/shared/huyuan/toolchains/
-#   AXERA: ax650n_bsp_sdk  |  AXCL: axcl_bsp_sdk
 
 TARGET=/data/shared/huyuan/toolchains
 mkdir -p $TARGET && cd $TARGET
@@ -10,10 +9,15 @@ if [ ! -d ax650n_bsp_sdk ]; then
   git clone https://github.com/AXERA-TECH/ax650n_bsp_sdk.git --depth=1
 fi
 
+if [ ! -d ax620e_bsp_sdk ]; then
+  echo "Downloading ax620e_bsp_sdk..."
+  git clone https://github.com/AXERA-TECH/ax620e_bsp_sdk.git --depth=1
+fi
+
 if [ ! -d axcl_bsp_sdk ]; then
   echo "Downloading axcl_bsp_sdk..."
   git clone https://github.com/Abandon-ht/axcl_bsp_sdk.git --depth=1
 fi
 
 echo "Done. SDKs at $TARGET/"
-ls -d ax650n_bsp_sdk axcl_bsp_sdk 2>/dev/null
+ls -d ax650n_bsp_sdk ax620e_bsp_sdk axcl_bsp_sdk 2>/dev/null

cpp/src/EngineWrapper_axera.cpp

@@ -9,7 +9,9 @@
 
 #include <cstdlib>
 
+#if !defined(AX630C)
 static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "10.8T"};
+#endif
 
 // NPU type enum
 typedef enum axNPU_TYPE_E {
@@ -21,6 +23,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2  = (1 << 4)
 } AX_NPU_TYPE_E;
 
+#if !defined(AX630C)
 static AX_S32 CheckModelVNpu(const std::string &strModel,
                               const AX_ENGINE_MODEL_TYPE_T &eModelType,
                               const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
@@ -57,6 +60,7 @@ static AX_S32 CheckModelVNpu(const std::string &strModel,
     }
     return ret;
 }
+#endif
 
 EngineWrapper::EngineWrapper()
     : m_hasInit(false), m_handle(nullptr), m_io_info(nullptr),
@@ -99,7 +103,8 @@ int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType, const char*
         }
     };
 
-    // 1.1 Get Model Type
+    // 1.1 Get Model Type & Check VNPU (AX650 only)
+#if !defined(AX630C)
     AX_ENGINE_MODEL_TYPE_T eModelType = AX_ENGINE_MODEL_TYPE0;
     AX_S32 ret = AX_ENGINE_GetModelType(pModelBufferVirAddr, nModelBufferSize, &eModelType);
     if (0 != ret || eModelType >= AX_ENGINE_MODEL_TYPE_BUTT) {
@@ -108,7 +113,6 @@ int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType, const char*
         return -1;
     }
 
-    // 1.2 Check VNPU
     AX_U32 nNpuSet = 0;
     ret = CheckModelVNpu(strModelPath, eModelType, nNpuType, nNpuSet);
     if (0 != ret) {
@@ -116,6 +120,10 @@ int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType, const char*
         freeModelBuffer();
         return -1;
     }
+#endif
+#if defined(AX630C)
+    AX_S32 ret = 0;
+#endif
 
     // 2. Create handle
     AX_ENGINE_HANDLE handle = nullptr;

cpp/src/vocoder.cpp

@@ -36,16 +36,26 @@ int Vocoder::Init(const Config& cfg) {
     }
     m_ifft_cfg = m_fft_cfg;
 
-    // Load vocos_full.axmodel
+    // Load backbone (or full) model
     m_session = std::make_unique<EngineWrapper>();
     if (m_session->Init(cfg.model_path.c_str(), 0, cfg.axclConfig) != 0) {
         printf("Failed to load vocoder model: %s\n", cfg.model_path.c_str());
         return -1;
     }
 
+    // Load head_linear model (optional, for split mode)
+    if (!cfg.head_model_path.empty()) {
+        m_session_head = std::make_unique<EngineWrapper>();
+        if (m_session_head->Init(cfg.head_model_path.c_str(), 0, cfg.axclConfig) != 0) {
+            printf("Failed to load vocoder head model: %s\n", cfg.head_model_path.c_str());
+            return -1;
+        }
+    }
+
     m_has_init = true;
-    printf("Vocoder initialized: n_fft=%d, hop=%d, n_mels=%d\n",
-           cfg.n_fft, cfg.hop_length, cfg.n_mels);
+    printf("Vocoder initialized: n_fft=%d, hop=%d, n_mels=%d%s\n",
+           cfg.n_fft, cfg.hop_length, cfg.n_mels,
+           m_session_head ? " (split mode)" : "");
     return 0;
 }
 
@@ -77,27 +87,54 @@ int Vocoder::Decode(const std::vector<float>& mel, int T, float feat_scale, std:
         }
     }
 
-    // 2. Run vocos_full.axmodel on NPU
-    int input_idx = m_session->GetInputIndex("mel");
-    if (input_idx < 0) input_idx = 0;
-    m_session->SetInput(mel_padded.data(), input_idx);
-
-    if (m_session->RunSync() != 0) {
-        printf("Vocoder NPU inference failed\n");
-        return -1;
+    // 2. Run NPU inference
+    std::vector<float> real_spec, imag_spec;
+    int spec_size;
+
+    if (m_session_head) {
+        // Split mode: backbone → features → head_linear → (real, imag)
+        int in_idx = m_session->GetInputIndex("mel");
+        if (in_idx < 0) in_idx = 0;
+        m_session->SetInput(mel_padded.data(), in_idx);
+        if (m_session->RunSync() != 0) { printf("Vocoder backbone failed\n"); return -1; }
+
+        int feat_idx = m_session->GetOutputIndex("features");
+        if (feat_idx < 0) feat_idx = 0;
+        std::vector<float> features(m_session->GetOutputSize(feat_idx) / sizeof(float));
+        m_session->GetOutput(features.data(), feat_idx);
+
+        int h_in = m_session_head->GetInputIndex("features");
+        if (h_in < 0) h_in = 0;
+        m_session_head->SetInput(features.data(), h_in);
+        if (m_session_head->RunSync() != 0) { printf("Vocoder head failed\n"); return -1; }
+
+        int real_idx = m_session_head->GetOutputIndex("real");
+        int imag_idx = m_session_head->GetOutputIndex("imag");
+        if (real_idx < 0) real_idx = 0;
+        if (imag_idx < 0) imag_idx = 1;
+        spec_size = m_session_head->GetOutputSize(real_idx);
+        real_spec.resize(spec_size / sizeof(float));
+        imag_spec.resize(spec_size / sizeof(float));
+        m_session_head->GetOutput(real_spec.data(), real_idx);
+        m_session_head->GetOutput(imag_spec.data(), imag_idx);
+    } else {
+        // Full model: mel → (real, imag)
+        int in_idx = m_session->GetInputIndex("mel");
+        if (in_idx < 0) in_idx = 0;
+        m_session->SetInput(mel_padded.data(), in_idx);
+        if (m_session->RunSync() != 0) { printf("Vocoder NPU inference failed\n"); return -1; }
+
+        int real_idx = m_session->GetOutputIndex("real");
+        int imag_idx = m_session->GetOutputIndex("imag");
+        if (real_idx < 0) real_idx = 0;
+        if (imag_idx < 0) imag_idx = 1;
+        spec_size = m_session->GetOutputSize(real_idx);
+        real_spec.resize(spec_size / sizeof(float));
+        imag_spec.resize(spec_size / sizeof(float));
+        m_session->GetOutput(real_spec.data(), real_idx);
+        m_session->GetOutput(imag_spec.data(), imag_idx);
     }
 
-    int real_idx = m_session->GetOutputIndex("real");
-    int imag_idx = m_session->GetOutputIndex("imag");
-    if (real_idx < 0) real_idx = 0;
-    if (imag_idx < 0) imag_idx = 1;
-
-    int spec_size = m_session->GetOutputSize(real_idx);
-    std::vector<float> real_spec(spec_size / sizeof(float));
-    std::vector<float> imag_spec(spec_size / sizeof(float));
-    m_session->GetOutput(real_spec.data(), real_idx);
-    m_session->GetOutput(imag_spec.data(), imag_idx);
-
     // 3. kissfft IRFFT + window + overlap-add
     int out_len = (T - 1) * hop + n_fft;
     audio.assign(out_len, 0.0f);

cpp/src/vocoder.hpp

@@ -16,11 +16,11 @@
 class Vocoder {
 public:
     struct Config {
-        std::string model_path;
-        const char* axclConfig = nullptr;   // AXCL only: path to axcl.json. Ignored on AXERA.
+        std::string model_path;         // full model (or backbone in split mode)
+        std::string head_model_path;    // head_linear model (optional, for split mode)
+        const char* axclConfig = nullptr;
         int n_fft = 1024;
         int hop_length = 256;
-        int win_length = 1024;
         int n_mels = 100;
         int sample_rate = 24000;
     };
@@ -48,6 +48,7 @@ private:
     Config m_cfg;
     bool m_has_init;
     std::unique_ptr<EngineWrapper> m_session;
+    std::unique_ptr<EngineWrapper> m_session_head;  // optional, for split mode
 
     // kissfft configs (shared forward/inverse twiddles)
     kiss_fft_state* m_fft_cfg = nullptr;

cpp/src/zipvoice_engine.cpp

@@ -159,6 +159,7 @@ int ZipVoiceEngine::LoadConfig(const std::string& model_dir) {
         m_config.hop_length = json_get_int(json, "hop_length", 256);
         m_config.num_step = json_get_int(json, "num_step", 10);
         m_config.t_shift = json_get_float(json, "t_shift", 0.5f);
+        m_config.guidance_scale = json_get_float(json, "guidance_scale", 1.0f);
     }
     m_config.model_dir = model_dir;
     return 0;
@@ -501,6 +502,7 @@ int ZipVoiceEngine::Sample(const std::vector<int32_t>& cat_tokens,
     std::copy(prompt_features.begin(),
               prompt_features.begin() + prompt_features_len * feat_dim,
               speech_cond_padded.begin());
+    // DEBUG
 
     // padding_mask: [1, seq_len] as uint8_t (bool), NOT float32!
     // Python: np.zeros((1, seq_len), dtype=np.bool_)

cpp/src/zipvoice_engine.hpp

@@ -39,6 +39,7 @@ public:
         int hop_length = 256;
         int num_step = 10;
         float t_shift = 0.5f;
+        float guidance_scale = 1.0f;
     };
 
     struct ModelInfo {

cpp/vocoder/vocos_full_ax630c.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b704aa370e7875099615ebbb2772c24e7d97529df81f11b713ca04e4261903e
+size 16208038

cpp/zipvoice.cpp

@@ -623,6 +623,8 @@ int main(int argc, char** argv) {
                          false, "");
     cmd.add<std::string>("vocoder-model", 0, "Path to vocos_full.axmodel for C++ vocoder. When set, all-C++ pipeline, no Python.",
                          false, "");
+    cmd.add<std::string>("vocoder-head-model", 0, "AX630C: path to head_linear axmodel (split mode, optional)",
+                         false, "");
     cmd.add<std::string>("axcl-config", 0, "AXCL only: path to axcl.json config file",
                          false, "/usr/local/axcl/axcl.json");
     cmd.add<int>("device-index", 0, "AXCL only: device index (0=first card)", false, 0);
@@ -636,7 +638,7 @@ int main(int argc, char** argv) {
     cmd.add<int>("max-feat-len", 0, "Max feature sequence length", false, 1024);
     cmd.add<int>("num-step", 0, "Number of flow-matching steps", false, 10);
     cmd.add<float>("speed", 0, "Speech speed factor", false, 1.0f);
-    cmd.add<float>("guidance-scale", 0, "CFG guidance scale", false, 1.0f);
+    cmd.add<float>("guidance-scale", 0, "CFG guidance scale (0=use config)", false, 0.0f);
     cmd.add<float>("t-shift", 0, "Time shift for flow scheduler", false, 0.5f);
     cmd.add<float>("feat-scale", 0, "Feature scaling factor", false, 0.1f);
     cmd.add<float>("target-rms", 0, "Target RMS for audio normalization", false, 0.1f);
@@ -658,6 +660,7 @@ int main(int argc, char** argv) {
     auto output_feat  = cmd.get<std::string>("output-feat");
     auto repo_dir     = cmd.get<std::string>("repo-dir");
     auto vocoder_model = cmd.get<std::string>("vocoder-model");
+    auto vocoder_head_model = cmd.get<std::string>("vocoder-head-model");
     auto axcl_config   = cmd.get<std::string>("axcl-config");
     int  device_index  = cmd.get<int>("device-index");
     auto cat_tokens_file = cmd.get<std::string>("cat-tokens-file");
@@ -871,6 +874,7 @@ int main(int argc, char** argv) {
         double t_vocoder_load = get_current_time_ms();
         Vocoder::Config vcfg;
         vcfg.model_path = vocoder_model;
+        vcfg.head_model_path = vocoder_head_model;
         vcfg.axclConfig = axcl_config.c_str();
         if (vocoder.Init(vcfg) != 0) {
             printf("ERROR: Failed to load vocoder model\n");
@@ -972,6 +976,7 @@ int main(int argc, char** argv) {
     ZipVoiceEngine::Timing timing;
     std::vector<float> output_features;
 
+    if (guidance_scale == 0.0f) guidance_scale = engine.GetConfig().guidance_scale;
     if (engine.Sample(cat_tokens,
                       prompt_tokens_len,
                       text_tokens_len,

infer_zipvoice_axera.py

@@ -58,6 +58,7 @@ def parse_args() -> argparse.Namespace:
         help="Allow raw features_len up to this ratio of max_feat_len before splitting",
     )
     p.add_argument("--silence-ms", type=int, default=140)
+    p.add_argument("--vocoder-model", default=None, help="Path to vocos axmodel. If set, use axmodel vocoder instead of PyTorch.")
     return p.parse_args()
 
 
@@ -158,7 +159,13 @@ def main() -> None:
         t_shift=args.t_shift,
     )
     logging.debug("Loading vocoder...")
-    vocoder = load_vocoder(repo_dir)
+    if args.vocoder_model:
+        from scripts.common_infer import load_axmodel_vocoder, axmodel_vocoder_decode
+        vocoder = load_axmodel_vocoder(args.vocoder_model)
+        vocoder_is_axmodel = True
+    else:
+        vocoder = load_vocoder(repo_dir)
+        vocoder_is_axmodel = False
     logging.info("模型加载完成")
 
     audios: list[np.ndarray] = []
@@ -183,13 +190,16 @@ def main() -> None:
             seed=args.seed + index - 1,
         )
 
-        audio = vocoder_decode_loaded(
-            vocoder,
-            pred_features,
-            feat_scale=args.feat_scale,
-            target_rms=args.target_rms,
-            prompt_rms=prompt_rms,
-        )
+        if vocoder_is_axmodel:
+            from scripts.common_infer import axmodel_vocoder_decode
+            audio = axmodel_vocoder_decode(
+                vocoder, pred_features,
+                feat_scale=args.feat_scale, target_rms=args.target_rms, prompt_rms=prompt_rms)
+        else:
+            audio = vocoder_decode_loaded(
+                vocoder, pred_features,
+                feat_scale=args.feat_scale, target_rms=args.target_rms,
+                prompt_rms=prompt_rms)
         segment_sec = time.perf_counter() - t_segment_start
         audio_sec = len(audio) / 24000
 

models/zipvoice_distill_ax630C/decoder4_split_manifest.json

@@ -1,68 +1,104 @@
 {
-    "version": 1,
-    "model_type": "zipvoice_distill",
-    "encoder": {
-        "name": "encoder_core",
-        "file": "encoder.axmodel",
-        "inputs": [
-            "cat_tokens"
-        ],
-        "outputs": [
-            "encoded"
-        ]
-    },
-    "decoder_parts": [
-        {
-            "name": "fm_decoder_part0",
-            "file": "decoder_part0.axmodel",
-            "inputs": [
-                "t",
-                "x",
-                "text_condition",
-                "speech_condition",
-                "guidance_scale",
-                "padding_mask"
-            ],
-            "outputs": [
-                "decoder_hidden_p0",
-                "time_hidden"
-            ]
-        },
-        {
-            "name": "fm_decoder_part1",
-            "file": "decoder_part1.axmodel",
-            "inputs": [
-                "decoder_hidden_p0",
-                "time_hidden",
-                "padding_mask"
-            ],
-            "outputs": [
-                "decoder_hidden_p1"
-            ]
-        },
-        {
-            "name": "fm_decoder_part2",
-            "file": "decoder_part2.axmodel",
-            "inputs": [
-                "decoder_hidden_p1",
-                "time_hidden",
-                "padding_mask"
-            ],
-            "outputs": [
-                "decoder_hidden_p2"
-            ]
-        },
-        {
-            "name": "fm_decoder_part3",
-            "file": "decoder_part3.axmodel",
-            "inputs": [
-                "decoder_hidden_p2",
-                "time_hidden",
-                "padding_mask"
-            ],
-            "outputs": [
-                "v"
-            ]
-        }
+  "version": 1,
+  "model_type": "zipvoice_distill",
+  "encoder": {
+    "name": "encoder_core",
+    "file": "encoder.axmodel",
+    "inputs": [
+      "cat_tokens"
+    ],
+    "outputs": [
+      "encoded"
     ]
-}
+  },
+  "decoder_parts": [
+    {
+      "name": "fm_decoder_part0",
+      "file": "decoder_part0.axmodel",
+      "inputs": [
+        "t",
+        "x",
+        "text_condition",
+        "speech_condition",
+        "guidance_scale",
+        "padding_mask"
+      ],
+      "outputs": [
+        "decoder_hidden_p0",
+        "time_hidden"
+      ]
+    },
+    {
+      "name": "fm_decoder_part1",
+      "file": "decoder_part1.axmodel",
+      "inputs": [
+        "decoder_hidden_p0",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "decoder_hidden_p1"
+      ]
+    },
+    {
+      "name": "fm_decoder_part2",
+      "file": "decoder_part2.axmodel",
+      "inputs": [
+        "decoder_hidden_p1",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "decoder_hidden_p2"
+      ]
+    },
+    {
+      "name": "fm_decoder_part3_0",
+      "file": "decoder_part3_0.axmodel",
+      "inputs": [
+        "decoder_hidden_p2",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "/fm_decoder/4/0/bypass/Add_output_0"
+      ]
+    },
+    {
+      "name": "fm_decoder_part3_1",
+      "file": "decoder_part3_1.axmodel",
+      "inputs": [
+        "/fm_decoder/4/0/bypass/Add_output_0",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "/fm_decoder/4/1/bypass/Add_output_0"
+      ]
+    },
+    {
+      "name": "fm_decoder_part3_2",
+      "file": "decoder_part3_2.axmodel",
+      "inputs": [
+        "/fm_decoder/4/1/bypass/Add_output_0",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "/fm_decoder/4/2/bypass/Add_output_0"
+      ]
+    },
+    {
+      "name": "fm_decoder_part3_3",
+      "file": "decoder_part3_3.axmodel",
+      "inputs": [
+        "/fm_decoder/4/2/bypass/Add_output_0",
+        "time_hidden",
+        "padding_mask"
+      ],
+      "outputs": [
+        "v"
+      ]
+    }
+  ]
+}

models/zipvoice_distill_ax630C/decoder_part0.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f0bd9974d94f349c0001d624d3c2c7140295abbe9c12c0db3556d5759d845a8
-size 42595736
+oid sha256:2113f394363a575904669017de685b4d8d73693736a285a4ce34782eba1a4531
+size 38600656

models/zipvoice_distill_ax630C/decoder_part1.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cfe296e53b7b8f3842aea1bbc60e48f946c0b5f0ae57c639faa4d31d75ff9c5e
-size 31881170
+oid sha256:6afd89c302f971b10677f27093bf60e7cb8f81ec191cde5c43781954976142be
+size 32510786

models/zipvoice_distill_ax630C/decoder_part2.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e95d65c96ba38c7a3cee6ecc558567a4d79e1803ba427749442690afc3b99da9
-size 35101628
+oid sha256:c883008c38b1ee727b8a68989dd4861d368a1cb593f4548e44fc6200dee89e6d
+size 35030020

models/zipvoice_distill_ax630C/decoder_part3_0.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b4d6a68bbc3286ed14bbe1d7f783bcc8cd559fe709151a99079d6558c1ce7d9
+size 10038370

models/zipvoice_distill_ax630C/decoder_part3_1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:744d3121d74670fec1f29f9a7ce5d158cb215e9b822464ad123fe8daeb8247f1
+size 10038193

models/zipvoice_distill_ax630C/decoder_part3_2.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:719c53fb3e4e0949b893c25814f705d33b12cd442fb84434d00dc3867c846ac1
+size 10038201

models/zipvoice_distill_ax630C/decoder_part3_3.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c44654f65f1398e611b02d9da2a23e931008ffa9ee6a53ebf027fd0da04bd17
+size 10004459

models/zipvoice_distill_ax630C/encoder.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ffbd78ded4a05cf472f78a66e34d9c4dac4c5525f6a7b1d2cc5dbb4d96e7538f
-size 8394155
+oid sha256:7193cbc302f3063522194b8287e1adfc250b83ab67c38c064bba74bef3eacb94
+size 7659503

run_ax630c.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+# ZipVoice AX630C 板端一键运行
+# 用法: bash run_ax630c.sh [zh|en] [sentence|paragraph]
+
+MODEL="${1:-zh}"
+MODE="${2:-sentence}"
+
+BIN="cpp/install/ax630c/zipvoice"
+MODEL_DIR="./models/zipvoice_distill_ax630C"
+TOKEN="./resources/zipvoice_hf/zipvoice/tokens.txt"
+VOCODER="./cpp/vocoder/vocos_full_ax630c.axmodel"
+
+cd "$(dirname "$0")"
+
+if [ "$MODEL" = "en" ]; then
+    PWAV="./assets/moss_prompts/en_4_4p5s.wav"
+    PTEXT="This is almost twice the current industry production level per train."
+    DTEXT="This morning, a small train left the station, carrying sleepy passengers toward a bright coastal town."
+    TFILE="./assets/paragraphs/en_scavenger.txt"
+    REPO="--repo-dir ."
+else
+    PWAV="./assets/moss_prompts/zh_1_4p5s.wav"
+    PTEXT="不管怎么样我和汤姆还是要感谢贝尔卡金的援手"
+    DTEXT="今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。"
+    TFILE="./assets/paragraphs/zh_ginkgo.txt"
+    REPO=""
+fi
+
+OUT="output_ax630c_${MODEL}_${MODE}.wav"
+
+echo "=========================================="
+echo " ZipVoice AX630C"
+echo "=========================================="
+echo " MODEL:  $MODEL_DIR"
+echo " LANG:   $MODEL  MODE: $MODE"
+echo " OUTPUT: $OUT"
+echo "=========================================="
+
+if [ "$MODE" = "paragraph" ]; then
+    exec $BIN \
+        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
+        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
+        --text-file "$TFILE" \
+        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
+else
+    exec $BIN \
+        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
+        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
+        --text "$DTEXT" \
+        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
+fi

scripts/common_infer.py

@@ -66,3 +66,74 @@ def vocoder_decode_loaded(
     if prompt_rms < target_rms:
         wav = wav * prompt_rms / target_rms
     return wav.squeeze().cpu().numpy()
+
+
+def load_axmodel_vocoder(model_path: str):
+    """Load vocoder axmodel for inference."""
+    import axengine as axe
+
+    session = axe.InferenceSession(model_path)
+    return session
+
+
+def axmodel_vocoder_decode(session, features: np.ndarray, feat_scale: float,
+                            target_rms: float, prompt_rms: float) -> np.ndarray:
+    """Decode mel features to audio using axmodel vocoder + IRFFT."""
+    import math
+
+    # features shape: (1, T, 100) or (T, 100), squeeze batch dim
+    features = np.squeeze(features)
+    if features.ndim == 2:
+        features = features.T  # (T, 100) → (100, T)
+    T = features.shape[1]
+    n_mels = features.shape[0]
+    n_fft, hop = 1024, 256
+    T_model = 620
+
+    # Undo feat_scale and pad to [1, n_mels, T_model]
+    inv_scale = 1.0 / feat_scale if feat_scale != 0 else 1.0
+    mel_input = np.zeros((1, n_mels, T_model), dtype=np.float32)
+    t_actual = min(T, T_model)
+    for t in range(t_actual):
+        mel_input[0, :, t] = features[:, t] * inv_scale
+
+    # Run axmodel: mel → (real, imag)
+    real, imag = session.run(None, {"mel": mel_input})
+    real, imag = real.squeeze(0), imag.squeeze(0)  # [T_model, n_freqs]
+
+    # IRFFT + overlap-add (matching C++ vocoder)
+    n_freqs = n_fft // 2 + 1
+    window = 0.5 * (1.0 - np.cos(2.0 * math.pi * np.arange(n_fft) / (n_fft - 1)))
+    window_sq = window ** 2
+    out_len = (T - 1) * hop + n_fft
+    audio = np.zeros(out_len, dtype=np.float32)
+    envelope = np.zeros(out_len, dtype=np.float32)
+
+    for t_idx in range(T):
+        spec = np.zeros(n_fft, dtype=np.complex64)
+        spec[0] = real[t_idx, 0]
+        for k in range(1, n_freqs - 1):
+            spec[k] = real[t_idx, k] + 1j * imag[t_idx, k]
+            spec[n_fft - k] = real[t_idx, k] - 1j * imag[t_idx, k]
+        spec[n_freqs - 1] = real[t_idx, n_freqs - 1]
+        ifft_out = np.fft.irfft(spec, n=n_fft).real
+
+        pos = t_idx * hop
+        for n in range(n_fft):
+            p = pos + n
+            if p < out_len:
+                audio[p] += ifft_out[n] * window[n]
+                envelope[p] += window_sq[n]
+
+    audio /= np.maximum(envelope, 1e-10)
+    pad = n_fft // 2
+    audio = audio[pad:out_len - pad].astype(np.float32)
+
+    # RMS normalize (numpy version)
+    rms = np.sqrt(np.mean(audio ** 2))
+    if rms < target_rms and rms > 1e-10:
+        audio = audio * (target_rms / rms)
+    if prompt_rms < target_rms:
+        audio = audio * (prompt_rms / target_rms)
+    return audio
+

scripts/zipvoice_decoder4_runtime.py

@@ -276,6 +276,7 @@ class Decoder4ZipVoiceBoardRuntime:
         encoded = self.run_encoder(cat_tokens)
         t_enc = time.perf_counter() - t_start
         logging.debug("  encoder: %.3f s (output shape=%s)", t_enc, encoded.shape)
+        # DEBUG: dump intermediates
 
         t_start = time.perf_counter()
         text_condition, features_len = self.duration_expand(
@@ -287,6 +288,7 @@ class Decoder4ZipVoiceBoardRuntime:
         )
         t_dur = time.perf_counter() - t_start
         logging.debug("  duration_expand: %.3f s (features_len=%d)", t_dur, features_len)
+        # DEBUG: dump
 
         seq_len = self.decoder_seq_len or self.max_feat_len
         if features_len > seq_len:

scripts/zipvoice_decoder4_runtime_encoder_onnx.py

@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Dict, List
+
+import numpy as np
+import onnxruntime as ort
+
+from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
+from scripts.zipvoice_runtime import AxeSession
+
+
+class _OrtSession:
+    """ONNX Runtime wrapper with the AxeSession interface used by split runtime."""
+
+    _TYPE_TO_DTYPE = {
+        "tensor(float)": np.float32,
+        "tensor(float32)": np.float32,
+        "tensor(double)": np.float64,
+        "tensor(int32)": np.int32,
+        "tensor(int64)": np.int64,
+        "tensor(uint8)": np.uint8,
+        "tensor(bool)": np.bool_,
+    }
+
+    def __init__(self, model_path: str | Path):
+        self.path = Path(model_path)
+        if not self.path.exists():
+            raise FileNotFoundError(f"ONNX model not found: {self.path}")
+        self._session = ort.InferenceSession(
+            str(self.path),
+            providers=["CPUExecutionProvider"],
+        )
+        self._inputs = self._session.get_inputs()
+        self._outputs = self._session.get_outputs()
+
+    @property
+    def input_names(self) -> List[str]:
+        return [item.name for item in self._inputs]
+
+    @property
+    def output_names(self) -> List[str]:
+        return [item.name for item in self._outputs]
+
+    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
+        info = next((item for item in self._inputs if item.name == name), None)
+        array = np.asarray(value)
+        if info is None:
+            return np.ascontiguousarray(array)
+
+        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
+        if dtype is not None:
+            array = array.astype(dtype, copy=False)
+
+        if list(info.shape) == [] and array.shape == (1,):
+            array = array.reshape(())
+
+        return np.ascontiguousarray(array)
+
+    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
+        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
+        outputs = self._session.run(None, feed)
+        return {name: value for name, value in zip(self.output_names, outputs)}
+
+
+class Decoder4ZipVoiceBoardRuntimeEncoderOnnx(Decoder4ZipVoiceBoardRuntime):
+    """Runs encoder with ONNX Runtime, all decoder parts with axmodel."""
+
+    def _load_models(self) -> None:
+        self.sessions = {}
+
+        encoder_name = self.encoder_info["name"]
+        encoder_path = self.models_dir / "encoder_core.onnx"
+        logging.info("encoder 使用 ONNX Runtime: %s", encoder_path)
+        self.sessions[encoder_name] = _OrtSession(encoder_path)
+
+        for info in self.decoder_parts:
+            name = info["name"]
+            path = self.models_dir / info["file"]
+            logging.debug("Loading %s from %s", name, path)
+            self.sessions[name] = AxeSession(path)
+
+        self.decoder_label = "decoder4(encoder_onnx+part0-3_axmodel)"
+        logging.debug("Loaded encoder ONNX + %d decoder axmodels", len(self.decoder_parts))

scripts/zipvoice_decoder4_runtime_part0_onnx.py

@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Dict, List
+
+import numpy as np
+import onnxruntime as ort
+
+from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
+from scripts.zipvoice_runtime import AxeSession
+
+
+class _OrtSession:
+    """ONNX Runtime wrapper with the small AxeSession interface used by the board runtime."""
+
+    _TYPE_TO_DTYPE = {
+        "tensor(float)": np.float32,
+        "tensor(float32)": np.float32,
+        "tensor(double)": np.float64,
+        "tensor(int32)": np.int32,
+        "tensor(int64)": np.int64,
+        "tensor(uint8)": np.uint8,
+        "tensor(bool)": np.bool_,
+    }
+
+    def __init__(self, model_path: str | Path):
+        self.path = Path(model_path)
+        if not self.path.exists():
+            raise FileNotFoundError(f"ONNX model not found: {self.path}")
+        self._session = ort.InferenceSession(
+            str(self.path),
+            providers=["CPUExecutionProvider"],
+        )
+        self._inputs = self._session.get_inputs()
+        self._outputs = self._session.get_outputs()
+
+    @property
+    def input_names(self) -> List[str]:
+        return [item.name for item in self._inputs]
+
+    @property
+    def output_names(self) -> List[str]:
+        return [item.name for item in self._outputs]
+
+    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
+        info = next((item for item in self._inputs if item.name == name), None)
+        array = np.asarray(value)
+        if info is None:
+            return np.ascontiguousarray(array)
+
+        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
+        if dtype is not None:
+            array = array.astype(dtype, copy=False)
+
+        # The exported part0 ONNX keeps t/guidance_scale as scalar inputs ([]),
+        # while the axmodel path feeds them as shape [1]. Normalize only for ONNX.
+        if list(info.shape) == [] and array.shape == (1,):
+            array = array.reshape(())
+
+        return np.ascontiguousarray(array)
+
+    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
+        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
+        outputs = self._session.run(None, feed)
+        return {name: value for name, value in zip(self.output_names, outputs)}
+
+
+class Decoder4ZipVoiceBoardRuntimePart0Onnx(Decoder4ZipVoiceBoardRuntime):
+    """Runs decoder part0 with ONNX Runtime, all other split models with axmodel."""
+
+    def _load_models(self) -> None:
+        self.sessions = {}
+
+        encoder_name = self.encoder_info["name"]
+        encoder_path = self.models_dir / self.encoder_info["file"]
+        logging.debug("Loading %s from %s", encoder_name, encoder_path)
+        self.sessions[encoder_name] = AxeSession(encoder_path)
+
+        for index, info in enumerate(self.decoder_parts):
+            name = info["name"]
+            if index == 0:
+                path = self.models_dir / "fm_decoder_part0.onnx"
+                logging.info("part0 使用 ONNX Runtime: %s", path)
+                self.sessions[name] = _OrtSession(path)
+            else:
+                path = self.models_dir / info["file"]
+                logging.debug("Loading %s from %s", name, path)
+                self.sessions[name] = AxeSession(path)
+
+        self.decoder_label = "decoder4(part0_onnx+part1-3_axmodel)"
+        logging.debug("Loaded encoder axmodel + part0 ONNX + %d decoder axmodels", len(self.decoder_parts) - 1)

scripts/zipvoice_decoder4_runtime_part3_onnx.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Dict, List
+
+import numpy as np
+import onnxruntime as ort
+
+from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
+from scripts.zipvoice_runtime import AxeSession
+
+
+class _OrtSession:
+    """ONNX Runtime wrapper with the AxeSession interface used by split runtime."""
+
+    _TYPE_TO_DTYPE = {
+        "tensor(float)": np.float32,
+        "tensor(float32)": np.float32,
+        "tensor(double)": np.float64,
+        "tensor(int32)": np.int32,
+        "tensor(int64)": np.int64,
+        "tensor(uint8)": np.uint8,
+        "tensor(bool)": np.bool_,
+    }
+
+    def __init__(self, model_path: str | Path):
+        self.path = Path(model_path)
+        if not self.path.exists():
+            raise FileNotFoundError(f"ONNX model not found: {self.path}")
+        self._session = ort.InferenceSession(
+            str(self.path),
+            providers=["CPUExecutionProvider"],
+        )
+        self._inputs = self._session.get_inputs()
+        self._outputs = self._session.get_outputs()
+
+    @property
+    def input_names(self) -> List[str]:
+        return [item.name for item in self._inputs]
+
+    @property
+    def output_names(self) -> List[str]:
+        return [item.name for item in self._outputs]
+
+    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
+        info = next((item for item in self._inputs if item.name == name), None)
+        array = np.asarray(value)
+        if info is None:
+            return np.ascontiguousarray(array)
+
+        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
+        if dtype is not None:
+            array = array.astype(dtype, copy=False)
+
+        if list(info.shape) == [] and array.shape == (1,):
+            array = array.reshape(())
+
+        return np.ascontiguousarray(array)
+
+    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
+        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
+        outputs = self._session.run(None, feed)
+        return {name: value for name, value in zip(self.output_names, outputs)}
+
+
+class Decoder4ZipVoiceBoardRuntimePart3Onnx(Decoder4ZipVoiceBoardRuntime):
+    """Runs decoder part3 with ONNX Runtime, encoder and part0-2 with axmodel."""
+
+    def _load_models(self) -> None:
+        self.sessions = {}
+
+        encoder_name = self.encoder_info["name"]
+        encoder_path = self.models_dir / self.encoder_info["file"]
+        logging.debug("Loading %s from %s", encoder_name, encoder_path)
+        self.sessions[encoder_name] = AxeSession(encoder_path)
+
+        last_index = len(self.decoder_parts) - 1
+        for index, info in enumerate(self.decoder_parts):
+            name = info["name"]
+            if index == last_index:
+                path = self.models_dir / "fm_decoder_part3.onnx"
+                logging.info("part3 使用 ONNX Runtime: %s", path)
+                self.sessions[name] = _OrtSession(path)
+            else:
+                path = self.models_dir / info["file"]
+                logging.debug("Loading %s from %s", name, path)
+                self.sessions[name] = AxeSession(path)
+
+        self.decoder_label = "decoder4(part0-2_axmodel+part3_onnx)"
+        logging.debug("Loaded encoder axmodel + %d decoder axmodels + part3 ONNX", last_index)