fix(audio): ignore playback tail before VAD

ql-application/threadx/interface/custom_app/vad_det.c

@@ -29,6 +29,9 @@
 // 在文件顶部添加状态变量
 static uint8_t speech_detect_count = 0;  // 连续语音帧计数器
 #define SPEECH_DETECT_THRESHOLD  20        // 连续检测到语音的阈值
+#define PLAYBACK_VAD_GUARD_MS     800       // 屏蔽播放结束后的尾音和音频通路切换噪声
+#define VAD_FRAME_BYTES           640       // 20ms @ 16kHz, 16-bit mono
+#define VAD_PREROLL_FRAMES        25        // 保留500ms预录音，避免吞掉用户首字
 uint8_t no_speech_timeout = 0;
 extern int g_mic_working;
 
@@ -57,6 +60,9 @@ static int64_t no_sound_start_time = 0;
 static int64_t cmd_wait_start_time = 0;
 static bool first_sound_detected = false;
 static int64_t recording_start_time = 0;  // 记录录音开始时间（首次检测到声音时）
+static uint8_t vad_preroll[VAD_PREROLL_FRAMES][VAD_FRAME_BYTES];
+static size_t vad_preroll_next = 0;
+static size_t vad_preroll_count = 0;
 
 static int vad_tts_playback_finished_status = 0;
 // 统计
@@ -101,6 +107,9 @@ void set_state(system_state_t new_state) {
                 no_sound_start_time = now;
                 first_sound_detected = false;
                 speech_frame_count = 0;
+				speech_detect_count = 0;
+				vad_preroll_next = 0;
+				vad_preroll_count = 0;
 				g_mic_working = 0;
                 break;
             case STATE_RECORDING_ACTIVE:
@@ -299,8 +308,8 @@ void vad_audio_process(void) {
     int64_t now = hal_os_up_time_ms();
     
     static size_t buffer_offset = 0;
-	uint8_t mic_buf[640]={0};
+	uint8_t mic_buf[VAD_FRAME_BYTES]={0};
-	int record_len = 640;
+	int record_len = VAD_FRAME_BYTES;
 	uint16 frame[320];
 	int index = 0, i = 0;
     
@@ -326,15 +335,19 @@ void vad_audio_process(void) {
 			//每次读取640字节音频数据
 			ql_audio_record_read(record_hd, mic_buf, &record_len); 
 
-			//如果检测到声音再去传输
+			if (g_state == STATE_RECORDING_WAIT_SOUND) {
-		//	if(first_sound_detected){
+				/* 首次开口确认前不向服务端发送静音。只在本地保留最近500ms，
-			//实际上每次编码需要输入1152字节的PCM数据
+				 * 确认开口时再回填，防止截掉用户首字。 */
-			if(hal_os_rb_write(ring_sender, (uint8*)mic_buf, record_len, 25)>0){
+				memcpy(vad_preroll[vad_preroll_next], mic_buf, VAD_FRAME_BYTES);
-
+				vad_preroll_next = (vad_preroll_next + 1) % VAD_PREROLL_FRAMES;
-				//quec_usbcdc_write((uint8*)mic_buf, record_len);
+				if (vad_preroll_count < VAD_PREROLL_FRAMES) {
-				ql_rtos_semaphore_release(sender_sem); //告诉线程去获取录音数据
+					vad_preroll_count++;
+				}
+			} else {
+				if(hal_os_rb_write(ring_sender, (uint8*)mic_buf, record_len, 25)>0){
+					ql_rtos_semaphore_release(sender_sem);
+				}
 			}
-		//	}
 		#if 1
 			for(index=0; index<320; index++)
 			{
@@ -347,6 +360,12 @@ void vad_audio_process(void) {
             
             // 首次声音检测
             if (!first_sound_detected) {
+				/* 播放刚结束时扬声器尾音和PA切换噪声容易被FVAD认为用户开口。 */
+				if ((now - no_sound_start_time) < PLAYBACK_VAD_GUARD_MS) {
+					speech_detect_count = 0;
+					buffer_offset = 0;
+					break;
+				}
                 if (current_vad == VAD_STATE_ACTIVE) {
                     speech_detect_count++;
                     printf("Speech frame count: %d\r\n", speech_detect_count);
@@ -363,8 +382,19 @@ void vad_audio_process(void) {
 						}
 						#endif
 						
-                        first_sound_detected = true;
+	                        first_sound_detected = true;
-                        set_state(STATE_RECORDING_ACTIVE);
+						{
+							size_t first = (vad_preroll_next + VAD_PREROLL_FRAMES - vad_preroll_count) % VAD_PREROLL_FRAMES;
+							size_t n;
+							for (n = 0; n < vad_preroll_count; n++) {
+								size_t frame_index = (first + n) % VAD_PREROLL_FRAMES;
+								if (hal_os_rb_write(ring_sender, vad_preroll[frame_index], VAD_FRAME_BYTES, 25) > 0) {
+									ql_rtos_semaphore_release(sender_sem);
+								}
+							}
+							vad_preroll_count = 0;
+						}
+	                        set_state(STATE_RECORDING_ACTIVE);
                         printf("First speech confirmed by FVAD (continuous %d frames)", SPEECH_DETECT_THRESHOLD);
                     }
                 } else {