Vocal　pattern　recognition　provides　an　objective　reference　for　subthreshold　depression　recognition　diagnosis　and　intervention　evaluation.　In　this　study,　we　used　different　speech　modalities　(reading　/a:/　tone,　text　reading,　picture　description,　free　interview)　and　emotional　stimuli　(positive,　neutral,　negative),　fused　four　types　of　speech　features　such　as　rhyme,　timbre,　spectrum,　and　resonance　peak,　extracted　16　feature　parameters　such　as　Mel-frequency　cepstrum　coefficient,　speed　of　sound,　fundamental　frequency,　and　resonance　peak,　established　a　subthreshold　depression　risk　prediction　model　using　random　forest,　and　compared　the　performance　with　other　classifiers.　The　results　showed　that　the　recognition　rate　of　picture　description　and　free　interview　before　fusing　features　was　higher　than　other　speech　modalities,　in　which　the　prediction　results　of　positive　stimuli　were　better　with　72.50%　and　67.39%　accuracy;　the　high　accuracy　rates　of　93.00%　and　85.00%　were　obtained　for　reading　/a:/　tone　and　free　interview　after　fusing　feature　layers,　respectively.　It　can　be　seen　that　the　phonetic　information　learned　by　the　model　after　fusing　features　contains　not　only　the　subject＇s　emotional　state　but　also　the　interrelationship　between　feature　types;　the　reading　/a:/　tone　and　free　interview　retain　more　vocal　tract　information,　where　the　reading　/a:/　tone　vocalization　is　persistent　and　the　sound　intensity　is　sustained,　and　the　free　interview　speech　volume　and　features　are　comprehensive　and　close　to　natural　speech,　which　are　informative　for　early　risk　prediction　of　subthreshold　depression.　©　2024　Journal　of　Fudan　University　(Natural　Science).　All　rights　reserved.