diff --git a/src/models/fire_red_vad/vad.rs b/src/models/fire_red_vad/vad.rs
index 0192e11..45ba9eb 100644
--- a/src/models/fire_red_vad/vad.rs
+++ b/src/models/fire_red_vad/vad.rs
@@ -9,7 +9,8 @@ use crate::{
         processor::{AudioFeat, VadPostprocessor},
     },
     utils::{
-        audio_utils::load_audio_with_resample_from_bytes, find_type_files, get_device,
+        audio_utils::{resample_audio_from_bytes, resample_audio_from_vec_f32},
+        find_type_files, get_device,
         tensor_utils::split_tensor_with_size,
     },
 };
@@ -84,13 +85,7 @@ impl FireRedVad {
         })
     }
 
-    pub fn detect_frame(&mut self, audio_bytes: Vec<u8>) -> Result<Option<VadFrameResult>> {
-        if !self.model_name.to_lowercase().contains("stream") {
-            return Err(anyhow!("only stream model support detect_frame"));
-        }
-        let audio_frame =
-            load_audio_with_resample_from_bytes(audio_bytes, &self.device, Some(16000), true)?
-                .squeeze(0)?;
+    pub fn detect_frame(&mut self, audio_frame: &Tensor) -> Result<Option<VadFrameResult>> {
         if audio_frame.dim(0)? < self.frame_length_sample {
             return Err(anyhow!(
                 "Expected {} samples, got {}",
@@ -98,7 +93,7 @@ impl FireRedVad {
                 audio_frame.dim(0)?
             ));
         }
-        let feats = self.audio_feat.extract(&audio_frame)?;
+        let feats = self.audio_feat.extract(audio_frame)?;
         let (probs, caches) = self
             .vad_model
             .forward(&feats.unsqueeze(0)?, self.caches.as_ref())?;
@@ -112,7 +107,7 @@ impl FireRedVad {
         if preds_sum as f32 > probs.dim(0)? as f32 * self.cfg.speech_threshold {
             Ok(Some(VadFrameResult {
                 is_speech: true,
-                orig_audio: Some(audio_frame),
+                orig_audio: Some(audio_frame.clone()),
                 kaldi_audio: Some(feats),
                 model_name: self.model_name.clone(),
                 mode: "speech".to_string(),
@@ -122,6 +117,36 @@ impl FireRedVad {
         }
     }
 
+    pub fn detect_frame_f32(
+        &mut self,
+        audio_vec_f32: Vec<f32>,
+        channels: usize,
+        orig_sr: Option<usize>,
+    ) -> Result<Option<VadFrameResult>> {
+        if !self.model_name.to_lowercase().contains("stream") {
+            return Err(anyhow!("only stream model support detect_frame"));
+        }
+        let audio_frame = resample_audio_from_vec_f32(
+            audio_vec_f32,
+            &self.device,
+            channels,
+            orig_sr,
+            Some(16000),
+            true,
+        )?
+        .squeeze(0)?;
+        self.detect_frame(&audio_frame)
+    }
+
+    pub fn detect_frame_bytes(&mut self, audio_bytes: Vec<u8>) -> Result<Option<VadFrameResult>> {
+        if !self.model_name.to_lowercase().contains("stream") {
+            return Err(anyhow!("only stream model support detect_frame"));
+        }
+        let audio_frame =
+            resample_audio_from_bytes(audio_bytes, &self.device, Some(16000), true)?.squeeze(0)?;
+        self.detect_frame(&audio_frame)
+    }
+
     pub fn detect_file(&self, audio_path: &str) -> Result<VadResult> {
         let (feats, dur) = self.audio_feat.extract_file(audio_path, &self.device)?;
         let probs = if feats.dim(0)? <= self.cfg.chunk_max_frame {
diff --git a/src/utils/audio_utils.rs b/src/utils/audio_utils.rs
index 169f4eb..f104831 100644
--- a/src/utils/audio_utils.rs
+++ b/src/utils/audio_utils.rs
@@ -587,7 +587,39 @@ pub fn load_audio(path: &str, device: &Device) -> Result<(Tensor, usize)> {
     load_audio_use_symphonia(audio_vec, false, device)
 }
 
-pub fn load_audio_with_resample_from_bytes(
+pub fn resample_audio_from_vec_f32(
+    audio_vec: Vec<f32>,
+    device: &Device,
+    channels: usize,
+    orig_sr: Option<usize>,
+    target_sample_rate: Option<usize>,
+    is_i16: bool,
+) -> Result<Tensor> {
+    let frame_len = audio_vec.len() / channels;
+    let audio = Tensor::new(&audio_vec[0..frame_len * channels], device)?;
+    let mut audio = if channels > 1 {
+        audio
+            .reshape((frame_len, channels))?
+            .mean_keepdim(1)?
+            .transpose(0, 1)?
+            .contiguous()?
+    } else {
+        audio.unsqueeze(0)?
+    };
+
+    if let Some(target_sample_rate) = target_sample_rate
+        && let Some(sr) = orig_sr
+        && target_sample_rate != sr
+    {
+        audio = resample_simple(&audio, sr as i64, target_sample_rate as i64)?;
+    }
+    if is_i16 {
+        audio = audio.affine(32768.0, 0.0)?;
+    }
+    Ok(audio)
+}
+
+pub fn resample_audio_from_bytes(
     audio_vec: Vec<u8>,
     device: &Device,
     target_sample_rate: Option<usize>,
@@ -612,7 +644,7 @@ pub fn load_audio_with_resample(
     // let audio_path = get_audio_path(path)?;
     // let (mut audio, sr) = load_audio_use_hound(audio_path, device)?;
     let audio_vec = get_audio_bytes_vec(path)?;
-    load_audio_with_resample_from_bytes(audio_vec, device, target_sample_rate, is_i16)
+    resample_audio_from_bytes(audio_vec, device, target_sample_rate, is_i16)
 }
 
 pub fn save_wav(audio: &Tensor, save_path: &str, sample_rate: u32) -> Result<()> {