diff --git a/core/http/endpoints/openai/realtime.go b/core/http/endpoints/openai/realtime.go
index 064658a7a7..d8546c30db 100644
--- a/core/http/endpoints/openai/realtime.go
+++ b/core/http/endpoints/openai/realtime.go
@@ -2122,7 +2122,9 @@ func vadScanWindowSec(sv *types.RealtimeSessionSemanticVad, silenceThreshold flo
 // decoded at 16 kHz (1.5x speed/pitch, garbage or empty transcript, then
 // input_not_committed on the follow-up response).
 func commitUtterance(ctx context.Context, utt []byte, session *Session, conv *Conversation, t Transport, slot *commitSlot) {
-	if rate := session.InputSampleRate; rate > 0 && rate != localSampleRate && len(utt) > 0 {
+	// Speech-to-speech sessions keep the audio at the session rate and label
+	// the transcription WAV with that rate (see s2sCommitTurn).
+	if rate := session.InputSampleRate; session.s2s == nil && rate > 0 && rate != localSampleRate && len(utt) > 0 {
 		utt = sound.Int16toBytesLE(sound.ResampleInt16(sound.BytesToInt16sLE(utt), rate, localSampleRate))
 	}
 	commitUtteranceWithTranscript(ctx, utt, nil, nil, "", session, conv, t, slot)
diff --git a/docs/content/features/openai-realtime.md b/docs/content/features/openai-realtime.md
index d628e5b893..b9a42859c2 100644
--- a/docs/content/features/openai-realtime.md
+++ b/docs/content/features/openai-realtime.md
@@ -589,7 +589,7 @@ ws://localhost:8080/v1/realtime?model=gpt-realtime
 
 Audio is sent and received as raw PCM in the WebSocket messages, following the OpenAI Realtime API protocol.
 
-Committed user audio reaches the transcription backend as 16 kHz mono WAV, whichever way the turn was committed: server-side VAD and client `input_audio_buffer.commit` both resample from the session input sample rate (for example 24 kHz) first.
+In the chained pipeline, committed user audio reaches the transcription backend as 16 kHz mono WAV, whichever way the turn was committed: server-side VAD and client `input_audio_buffer.commit` both resample from the session input sample rate (for example 24 kHz) first.
 
 ### WebRTC