diff --git a/controller/relay-audio.go b/controller/relay-audio.go
index 89a311a0..5b8898a7 100644
--- a/controller/relay-audio.go
+++ b/controller/relay-audio.go
@@ -39,41 +39,40 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode
 		}
 	}
 
-	preConsumedTokens := common.PreConsumedQuota
 	modelRatio := common.GetModelRatio(audioModel)
 	groupRatio := common.GetGroupRatio(group)
 	ratio := modelRatio * groupRatio
-	preConsumedQuota := int(float64(preConsumedTokens) * ratio)
+	var quota int
+	var preConsumedQuota int
+	switch relayMode {
+	case RelayModeAudioSpeech:
+		preConsumedQuota = int(float64(len(ttsRequest.Input)) * ratio)
+		quota = preConsumedQuota
+	default:
+		preConsumedQuota = int(float64(common.PreConsumedQuota) * ratio)
+	}
 	userQuota, err := model.CacheGetUserQuota(userId)
 	if err != nil {
 		return errorWrapper(err, "get_user_quota_failed", http.StatusInternalServerError)
 	}
 
-	quota := 0
 	// Check if user quota is enough
-	if relayMode == RelayModeAudioSpeech {
-		quota = int(float64(len(ttsRequest.Input)) * modelRatio * groupRatio)
-		if quota > userQuota {
-			return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden)
-		}
-	} else {
-		if userQuota-preConsumedQuota < 0 {
-			return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden)
-		}
-		err = model.CacheDecreaseUserQuota(userId, preConsumedQuota)
+	if userQuota-preConsumedQuota < 0 {
+		return errorWrapper(errors.New("user quota is not enough"), "insufficient_user_quota", http.StatusForbidden)
+	}
+	err = model.CacheDecreaseUserQuota(userId, preConsumedQuota)
+	if err != nil {
+		return errorWrapper(err, "decrease_user_quota_failed", http.StatusInternalServerError)
+	}
+	if userQuota > 100*preConsumedQuota {
+		// in this case, we do not pre-consume quota
+		// because the user has enough quota
+		preConsumedQuota = 0
+	}
+	if preConsumedQuota > 0 {
+		err := model.PreConsumeTokenQuota(tokenId, preConsumedQuota)
 		if err != nil {
-			return errorWrapper(err, "decrease_user_quota_failed", http.StatusInternalServerError)
-		}
-		if userQuota > 100*preConsumedQuota {
-			// in this case, we do not pre-consume quota
-			// because the user has enough quota
-			preConsumedQuota = 0
-		}
-		if preConsumedQuota > 0 {
-			err := model.PreConsumeTokenQuota(tokenId, preConsumedQuota)
-			if err != nil {
-				return errorWrapper(err, "pre_consume_token_quota_failed", http.StatusForbidden)
-			}
+			return errorWrapper(err, "pre_consume_token_quota_failed", http.StatusForbidden)
 		}
 	}
 
@@ -141,11 +140,7 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode
 		return errorWrapper(err, "close_request_body_failed", http.StatusInternalServerError)
 	}
 
-	if relayMode == RelayModeAudioSpeech {
-		defer func(ctx context.Context) {
-			go postConsumeQuota(ctx, tokenId, quota, userId, channelId, modelRatio, groupRatio, audioModel, tokenName)
-		}(c.Request.Context())
-	} else {
+	if relayMode != RelayModeAudioSpeech {
 		responseBody, err := io.ReadAll(resp.Body)
 		if err != nil {
 			return errorWrapper(err, "read_response_body_failed", http.StatusInternalServerError)
@@ -159,13 +154,29 @@ func relayAudioHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode
 		if err != nil {
 			return errorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError)
 		}
-		defer func(ctx context.Context) {
-			quota := countTokenText(whisperResponse.Text, audioModel)
-			quotaDelta := quota - preConsumedQuota
-			go postConsumeQuota(ctx, tokenId, quotaDelta, userId, channelId, modelRatio, groupRatio, audioModel, tokenName)
-		}(c.Request.Context())
+		quota = countTokenText(whisperResponse.Text, audioModel)
 		resp.Body = io.NopCloser(bytes.NewBuffer(responseBody))
 	}
+	if resp.StatusCode != http.StatusOK {
+		if preConsumedQuota > 0 {
+			// we need to roll back the pre-consumed quota
+			defer func(ctx context.Context) {
+				go func() {
+					// negative means add quota back for token & user
+					err := model.PostConsumeTokenQuota(tokenId, -preConsumedQuota)
+					if err != nil {
+						common.LogError(ctx, fmt.Sprintf("error rollback pre-consumed quota: %s", err.Error()))
+					}
+				}()
+			}(c.Request.Context())
+		}
+		return relayErrorHandler(resp)
+	}
+	quotaDelta := quota - preConsumedQuota
+	defer func(ctx context.Context) {
+		go postConsumeQuota(ctx, tokenId, quotaDelta, quota, userId, channelId, modelRatio, groupRatio, audioModel, tokenName)
+	}(c.Request.Context())
+
 	for k, v := range resp.Header {
 		c.Writer.Header().Set(k, v[0])
 	}
diff --git a/controller/relay-utils.go b/controller/relay-utils.go
index c7cd4766..391f28b4 100644
--- a/controller/relay-utils.go
+++ b/controller/relay-utils.go
@@ -195,8 +195,9 @@ func getFullRequestURL(baseURL string, requestURL string, channelType int) strin
 	return fullRequestURL
 }
 
-func postConsumeQuota(ctx context.Context, tokenId int, quota int, userId int, channelId int, modelRatio float64, groupRatio float64, modelName string, tokenName string) {
-	err := model.PostConsumeTokenQuota(tokenId, quota)
+func postConsumeQuota(ctx context.Context, tokenId int, quotaDelta int, totalQuota int, userId int, channelId int, modelRatio float64, groupRatio float64, modelName string, tokenName string) {
+	// quotaDelta is remaining quota to be consumed
+	err := model.PostConsumeTokenQuota(tokenId, quotaDelta)
 	if err != nil {
 		common.SysError("error consuming token remain quota: " + err.Error())
 	}
@@ -204,10 +205,14 @@ func postConsumeQuota(ctx context.Context, tokenId int, quota int, userId int, c
 	if err != nil {
 		common.SysError("error update user quota cache: " + err.Error())
 	}
-	if quota != 0 {
+	// totalQuota is total quota consumed
+	if totalQuota != 0 {
 		logContent := fmt.Sprintf("模型倍率 %.2f，分组倍率 %.2f", modelRatio, groupRatio)
-		model.RecordConsumeLog(ctx, userId, channelId, 0, 0, modelName, tokenName, quota, logContent)
-		model.UpdateUserUsedQuotaAndRequestCount(userId, quota)
-		model.UpdateChannelUsedQuota(channelId, quota)
+		model.RecordConsumeLog(ctx, userId, channelId, totalQuota, 0, modelName, tokenName, totalQuota, logContent)
+		model.UpdateUserUsedQuotaAndRequestCount(userId, totalQuota)
+		model.UpdateChannelUsedQuota(channelId, totalQuota)
+	}
+	if totalQuota <= 0 {
+		common.LogError(ctx, fmt.Sprintf("totalQuota consumed is %d, something is wrong", totalQuota))
 	}
 }