diff --git a/torchaudio/csrc/sox/effects_chain.cpp b/torchaudio/csrc/sox/effects_chain.cpp
index 696dd1adc86..9edf772de38 100644
--- a/torchaudio/csrc/sox/effects_chain.cpp
+++ b/torchaudio/csrc/sox/effects_chain.cpp
@@ -336,7 +336,7 @@ int fileobj_input_drain(sox_effect_t* effp, sox_sample_t* obuf, size_t* osamp) {
   const auto num_consumed = sf->tell_off;
   const auto num_remain = priv->buffer_size - num_consumed;
 
-  // 1.1. First, we fetch the data to see if there is data to fill the buffer
+  // 1.1. Fetch the data to see if there is data to fill the buffer
   uint64_t num_refill = 0;
   std::string chunk;
   if (num_consumed && !priv->read_finished) {
@@ -388,7 +388,8 @@ int fileobj_input_drain(sox_effect_t* effp, sox_sample_t* obuf, size_t* osamp) {
   // store the actual number read back to *osamp
   *osamp = sox_read(sf, obuf, *osamp);
 
-  return (!*osamp && priv->read_finished) ? SOX_EOF : SOX_SUCCESS;
+  // Decoding is finished when fileobject is exhausted and sox can no longer decode a sample.
+  return (priv->read_finished && !*osamp) ? SOX_EOF : SOX_SUCCESS;
 }
 
 int fileobj_output_flow(