Update README.md

.gitattributes

@@ -37,6 +37,11 @@ axllm filter=lfs diff=lfs merge=lfs -text
 models/SenseVoiceSmall/ax650/model-10-seconds.axmodel filter=lfs diff=lfs merge=lfs -text
 models/SenseVoiceSmall/ax650/model-5-seconds.axmodel filter=lfs diff=lfs merge=lfs -text
 models/SenseVoiceSmall/bin/sherpa-onnx-offline filter=lfs diff=lfs merge=lfs -text
+models/SenseVoiceSmall/lib/libonnxruntime.so filter=lfs diff=lfs merge=lfs -text
+models/SenseVoiceSmall/lib/libonnxruntime.so.1 filter=lfs diff=lfs merge=lfs -text
+models/SenseVoiceSmall/lib/libonnxruntime.so.1.23.2 filter=lfs diff=lfs merge=lfs -text
+models/SenseVoiceSmall/lib/libsherpa-onnx-c-api.so filter=lfs diff=lfs merge=lfs -text
+models/SenseVoiceSmall/lib/libsherpa-onnx-cxx-api.so filter=lfs diff=lfs merge=lfs -text
 models/Vad/ax_model/vad.axmodel filter=lfs diff=lfs merge=lfs -text
 models/Vad/wav/vad_example.wav filter=lfs diff=lfs merge=lfs -text
 models/melotts/decoder-zh.axmodel filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,6 +1,6 @@
 # Voice_Assistant.AXERA
 
-ax650板端实时语音对话助手demo。
+ax650板端实时语音对话助手demo
 
 ## 特征
 

asr_server.py

@@ -43,10 +43,12 @@ _MIN_DURATION_MS = 500
 
 
 def _model_for_duration(duration_seconds: float) -> Path:
-    model = _MODEL_ROOT / "ax650" / "model-5-seconds.axmodel"
-    if not model.exists():
-        raise FileNotFoundError(f"SenseVoice 5s model not found: {model}")
-    return model
+    for seconds in (5, 10, 15, 20, 25, 30):
+        if duration_seconds <= seconds:
+            model = _MODEL_ROOT / "ax650" / f"model-{seconds}-seconds.axmodel"
+            if model.exists():
+                return model
+    return _MODEL_ROOT / "ax650" / "model-10-seconds.axmodel"
 
 
 @app.get("/health")
@@ -138,6 +140,21 @@ def main():
 
     logger.info("ASR server ready (model_root=%s, provider=%s)", _MODEL_ROOT, _PROVIDER)
 
+    # warm up: pre-load the model so the first real request is fast
+    _warmup_path = os.path.join(tempfile.gettempdir(), "_asr_warmup.wav")
+    try:
+        _warmup_audio = np.zeros(int(_SAMPLE_RATE * 0.5), dtype=np.float32)  # 0.5s silence
+        sf.write(_warmup_path, _warmup_audio, _SAMPLE_RATE)
+        model_file = _MODEL_ROOT / "ax650" / "model-5-seconds.axmodel"
+        if model_file.exists():
+            cmd = [str(_SHERPA_BIN), f"--sense-voice-model={model_file}", f"--tokens={_TOKENS_FILE}", f"--provider={_PROVIDER}", _warmup_path]
+            t0 = time.time()
+            subprocess.run(cmd, capture_output=True, text=True, timeout=60)
+            logger.info("ASR warmup complete (%.1fs)", time.time() - t0)
+        os.remove(_warmup_path)
+    except Exception as e:
+        logger.warning("ASR warmup failed (non-fatal): %s", e)
+
     import uvicorn
     uvicorn.run(app, host=args.host, port=args.port)
 

models/SenseVoiceSmall/lib/cargs.h

@@ -0,0 +1,162 @@
+#pragma once
+
+/**
+ * This is a simple alternative cross-platform implementation of getopt, which
+ * is used to parse argument strings submitted to the executable (argc and argv
+ * which are received in the main function).
+ */
+
+#ifndef CAG_LIBRARY_H
+#define CAG_LIBRARY_H
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdio.h>
+
+#if defined(_WIN32) || defined(__CYGWIN__)
+#define CAG_EXPORT __declspec(dllexport)
+#define CAG_IMPORT __declspec(dllimport)
+#elif __GNUC__ >= 4
+#define CAG_EXPORT __attribute__((visibility("default")))
+#define CAG_IMPORT __attribute__((visibility("default")))
+#else
+#define CAG_EXPORT
+#define CAG_IMPORT
+#endif
+
+#if defined(CAG_SHARED)
+#if defined(CAG_EXPORTS)
+#define CAG_PUBLIC CAG_EXPORT
+#else
+#define CAG_PUBLIC CAG_IMPORT
+#endif
+#else
+#define CAG_PUBLIC
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * An option is used to describe a flag/argument option submitted when the
+ * program is run.
+ */
+typedef struct cag_option
+{
+  const char identifier;
+  const char *access_letters;
+  const char *access_name;
+  const char *value_name;
+  const char *description;
+} cag_option;
+
+/**
+ * A context is used to iterate over all options provided. It stores the parsing
+ * state.
+ */
+typedef struct cag_option_context
+{
+  const struct cag_option *options;
+  size_t option_count;
+  int argc;
+  char **argv;
+  int index;
+  int inner_index;
+  bool forced_end;
+  char identifier;
+  char *value;
+} cag_option_context;
+
+/**
+ * This is just a small macro which calculates the size of an array.
+ */
+#define CAG_ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
+
+/**
+ * @brief Prints all options to the terminal.
+ *
+ * This function prints all options to the terminal. This can be used to
+ * generate the output for a "--help" option.
+ *
+ * @param options The options which will be printed.
+ * @param option_count The option count which will be printed.
+ * @param destination The destination where the output will be printed.
+ */
+CAG_PUBLIC void cag_option_print(const cag_option *options, size_t option_count,
+  FILE *destination);
+
+/**
+ * @brief Prepare argument options context for parsing.
+ *
+ * This function prepares the context for iteration and initializes the context
+ * with the supplied options and arguments. After the context has been prepared,
+ * it can be used to fetch arguments from it.
+ *
+ * @param context The context which will be initialized.
+ * @param options The registered options which are available for the program.
+ * @param option_count The amount of options which are available for the
+ * program.
+ * @param argc The amount of arguments the user supplied in the main function.
+ * @param argv A pointer to the arguments of the main function.
+ */
+CAG_PUBLIC void cag_option_prepare(cag_option_context *context,
+  const cag_option *options, size_t option_count, int argc, char **argv);
+
+/**
+ * @brief Fetches an option from the argument list.
+ *
+ * This function fetches a single option from the argument list. The context
+ * will be moved to that item. Information can be extracted from the context
+ * after the item has been fetched.
+ * The arguments will be re-ordered, which means that non-option arguments will
+ * be moved to the end of the argument list. After all options have been
+ * fetched, all non-option arguments will be positioned after the index of
+ * the context.
+ *
+ * @param context The context from which we will fetch the option.
+ * @return Returns true if there was another option or false if the end is
+ * reached.
+ */
+CAG_PUBLIC bool cag_option_fetch(cag_option_context *context);
+
+/**
+ * @brief Gets the identifier of the option.
+ *
+ * This function gets the identifier of the option, which should be unique to
+ * this option and can be used to determine what kind of option this is.
+ *
+ * @param context The context from which the option was fetched.
+ * @return Returns the identifier of the option.
+ */
+CAG_PUBLIC char cag_option_get(const cag_option_context *context);
+
+/**
+ * @brief Gets the value from the option.
+ *
+ * This function gets the value from the option, if any. If the option does not
+ * contain a value, this function will return NULL.
+ *
+ * @param context The context from which the option was fetched.
+ * @return Returns a pointer to the value or NULL if there is no value.
+ */
+CAG_PUBLIC const char *cag_option_get_value(const cag_option_context *context);
+
+/**
+ * @brief Gets the current index of the context.
+ *
+ * This function gets the index within the argv arguments of the context. The
+ * context always points to the next item which it will inspect. This is
+ * particularly useful to inspect the original argument array, or to get
+ * non-option arguments after option fetching has finished.
+ *
+ * @param context The context from which the option was fetched.
+ * @return Returns the current index of the context.
+ */
+CAG_PUBLIC int cag_option_get_index(const cag_option_context *context);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+#endif

models/SenseVoiceSmall/lib/libonnxruntime.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b14fe799d56bb20f9175d707bc322d1395a38526ff9d0edd7eca1b7ab89d45cd
+size 18762289

models/SenseVoiceSmall/lib/libonnxruntime.so.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b14fe799d56bb20f9175d707bc322d1395a38526ff9d0edd7eca1b7ab89d45cd
+size 18762289

models/SenseVoiceSmall/lib/libonnxruntime.so.1.23.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b14fe799d56bb20f9175d707bc322d1395a38526ff9d0edd7eca1b7ab89d45cd
+size 18762289

models/SenseVoiceSmall/lib/libsherpa-onnx-c-api.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e0462c4f8f9ca1448ab5fb3526bc35cf06476e6c3f1975ece641c5fb5408e8b
+size 7078544

models/SenseVoiceSmall/lib/libsherpa-onnx-cxx-api.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47ce147df2081e3336c48a55ab1c3c99089edd61ecfeca67fb4d6a520297b061
+size 135192

models/SenseVoiceSmall/lib/pkgconfig/espeak-ng.pc

@@ -0,0 +1,11 @@
+prefix=/home/m5stack/Workspace/kaldi/sherpa-onnx/build-axera-linux-aarch64-ax650/install
+exec_prefix=/home/m5stack/Workspace/kaldi/sherpa-onnx/build-axera-linux-aarch64-ax650/install
+libdir=${exec_prefix}/lib
+includedir=${prefix}/include
+
+Name: espeak-ng
+Description: espeak-ng is a multi-lingual software speech synthesizer
+Version: 1.52.0.1
+Requires:
+Libs: -L${libdir} -lespeak-ng
+Cflags: -I${includedir}

stageflow/core/gateway.py

@@ -224,7 +224,7 @@ def create_app(config_path):
             return [
                 "python3", "asr_server.py",
                 "--model-root", str(asset_root / "SenseVoiceSmall"),
-                "--sherpa-bin", str(asset_root / "kokoro" / "build" / "bin" / "sherpa-onnx-offline"),
+                "--sherpa-bin", str(asset_root / "SenseVoiceSmall" / "bin" / "sherpa-onnx-offline"),
                 "--provider", cfg.get("provider", "axera"),
                 "--host", "0.0.0.0", "--port", "8010",
             ]

stageflow/units/speech_recognition.py

@@ -66,10 +66,12 @@ class SherpaSenseVoiceASR(ASR):
         self.sherpa_bin.chmod(0o755)
 
     def _model_for_duration(self, duration_seconds):
-        model = self.root / "ax650" / "model-5-seconds.axmodel"
-        if not model.exists():
-            raise FileNotFoundError(f"SenseVoice 5s model not found: {model}")
-        return model
+        for seconds in (5, 10, 15, 20, 25, 30):
+            if duration_seconds <= seconds:
+                model = self.root / "ax650" / f"model-{seconds}-seconds.axmodel"
+                if model.exists():
+                    return model
+        return self.root / "ax650" / "model-10-seconds.axmodel"
 
     @staticmethod
     def _clean_text(text):