This　study　introduces　a　novel　perspective　on　financial　fraud　detection　by　exploring　the　utility　of　managers＇　abnormal　tone.　To　mitigate　bias　in　indicator　selection,　we　implement　a　feature　selection　process　involving　a　comprehensive　set　of　301　indicators,　including　financial,　nonfinancial,　and　textual,　and　various　machine　learning　algorithms.　The　dataset　contains　6077　pairs　of　fraudulent　and　non-fraudulent　samples　in　China.　Our　findings　underscore　the　significance　of　abnormal　tone　in　fraud　detection,　establishing　it　as　a　prominent　factor　in　the　feature　selection　process.　The　accuracy　outcomes　from　eight　machine　learning　models　further　confirm　that　incorporating　abnormal　tone　can　enhance　fraud　detection　performance.