{"text": "你敢信吗?Github上跟Whisper相关的开源项目光星星过千的就有十几个,每个都说自己更快更强,但我全跑了一遍,真正能打的没几个。先说一个前提,我选Whisper的核心原因只有一个,多语言之耻。中文、英文、日文、韩文都得能识别。如果只做中文,我根本不会选Whisper,翻译S2它不香吗?但多语言这个需求目前只有Whisper能扛。先说几个不适合实时场景的。Whisper Finiton用LORAV条,训练成本低,还做了三端部署,工程完整度很高。但它是离线识别,先录完再转,流续识别用不了。Whisper GX速度快70倍,但峰值性能依赖TPU,普通GPU上优势没那么多。Buzz是桌面App,不适合后端部署。再说几个功能聚焦,但场景窄的。Faster Whisper用C-Translate 2加速,速度是原版4倍,内存更低,支持8位量化。行业里用得非常广,但它只是推理引擎,不提供流适输出,你得自己组合其他工具。YT Whisper只做YouTube字幕,场景太窄。Whisper Transparent只做时间戳,没有加速和部署。实时转录那一批呢?Whisper Mic场景最贴合,麦克风输入实时转录,但只支持英文。Whisper Real Time和Whisper Live思路类似,都是边录边识别,但都没有加速优化,中文延迟偏高。Whisper Realization倒是厉害,能识别谁在什么时候说了什么,但串了四五个模型,太重了。然后是两个重量级选手,Whisper X,综合评价最高。Large VR做70倍实时速度,不到8GB显存就能跑。关键是支持单词及时间戳,每个字精确对应到音频时间位置,做流逝展示和字幕对齐,这个能力太重要了,短期方案它是最稳的。另一个是Distil Whisper,把Large VR蒸留成小模型,速度快六倍,体积小一半,错误率只涨不到1%,性价比太炸了。但知名问题是,只发布了英文模型,其他语言得自己蒸留。门槛是有的,但走通了长期收益最大。所以方向很清晰,多语言需求选Whisper没毛病,但具体用哪个项目得看你的场景。短期上Whisper X最稳,长期自蒸留Distil Whisper收益最大。选项目不是选最好的,是选最合适的。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 9.540000000000001, "text": "你敢信吗?Github上跟Whisper相关的开源项目光星星过千的就有十几个,每个都说自己更快更强,但我全跑了一遍,真正能打的没几个。", "tokens": [50365, 2166, 40205, 17665, 14769, 30, 38, 355, 836, 5708, 9678, 2471, 271, 610, 15106, 28053, 1546, 18937, 47402, 10178, 117, 11386, 20690, 20682, 20682, 16866, 20787, 1546, 3111, 2412, 20145, 6336, 254, 7549, 11, 23664, 7549, 7182, 8090, 17645, 19002, 10251, 19002, 5702, 118, 11, 8395, 1654, 11319, 32585, 2289, 2257, 3330, 235, 11, 6303, 15789, 8225, 12467, 1546, 10062, 6336, 254, 7549, 1543, 50842], "temperature": 0, "avg_logprob": -0.08289278745651245, "compression_ratio": 1.29375, "no_speech_prob": 9.505813670929175e-12}, {"id": 1, "seek": 0, "start": 10.56, "end": 25.16, "text": "先说一个前提,我选Whisper的核心原因只有一个,多语言之耻。中文、英文、日文、韩文都得能识别。如果只做中文,我根本不会选Whisper,翻译S2它不香吗?但多语言这个需求目前只有Whisper能扛。", "tokens": [50893, 10108, 8090, 20182, 8945, 20949, 11, 1654, 2215, 231, 2471, 271, 610, 1546, 8987, 116, 7945, 42577, 35244, 20182, 11, 6392, 5233, 255, 12009, 9574, 4450, 119, 1543, 5975, 17174, 1231, 27869, 17174, 1231, 6890, 17174, 1231, 13665, 102, 17174, 7182, 5916, 8225, 5233, 228, 18453, 1543, 13119, 14003, 10907, 5975, 17174, 11, 1654, 31337, 8802, 47928, 2215, 231, 2471, 271, 610, 11, 42716, 5233, 239, 50, 17, 11284, 1960, 16017, 14769, 30, 8395, 6392, 5233, 255, 12009, 15368, 32535, 32718, 39004, 35244, 2471, 271, 610, 8225, 3416, 249, 1543, 51623], "temperature": 0, "avg_logprob": -0.08289278745651245, "compression_ratio": 1.29375, "no_speech_prob": 9.505813670929175e-12}, {"id": 2, "seek": 2516, "start": 25.16, "end": 33.76, "text": "先说几个不适合实时场景的。Whisper Finiton用LORAV条,训练成本低,还做了三端部署,工程完整度很高。", "tokens": [50365, 10108, 8090, 6336, 254, 7549, 1960, 2215, 224, 14245, 24726, 15729, 50255, 50218, 1546, 1543, 2471, 271, 610, 3773, 270, 266, 9254, 43, 2483, 32, 53, 48837, 11, 7422, 255, 10115, 225, 11336, 8802, 41377, 11, 14852, 10907, 2289, 10960, 11957, 107, 13470, 16469, 110, 11, 23323, 29649, 14128, 27662, 13127, 4563, 12979, 1543, 50795], "temperature": 0, "avg_logprob": -0.10234992672698666, "compression_ratio": 1.2086720867208671, "no_speech_prob": 1.2707303959080818e-11}, {"id": 3, "seek": 2516, "start": 34.0, "end": 43.04, "text": "但它是离线识别,先录完再转,流续识别用不了。Whisper GX速度快70倍,但峰值性能依赖TPU,普通GPU上优势没那么多。", "tokens": [50807, 8395, 11284, 1541, 18614, 119, 16853, 123, 5233, 228, 18453, 11, 10108, 7391, 243, 14128, 8623, 17819, 105, 11, 27854, 10115, 255, 5233, 228, 18453, 9254, 47225, 1543, 2471, 271, 610, 460, 55, 31217, 13127, 10251, 5867, 35477, 11, 8395, 26555, 108, 40242, 21686, 8225, 3254, 251, 5266, 244, 51, 8115, 11, 29993, 19550, 38, 8115, 5708, 7384, 246, 5087, 123, 10062, 35693, 6392, 1543, 51259], "temperature": 0, "avg_logprob": -0.10234992672698666, "compression_ratio": 1.2086720867208671, "no_speech_prob": 1.2707303959080818e-11}, {"id": 4, "seek": 2516, "start": 43.72, "end": 45.74, "text": "Buzz是桌面App,不适合后端部署。", "tokens": [51293, 33, 16740, 1541, 162, 5475, 8833, 9132, 11, 1960, 2215, 224, 14245, 13547, 11957, 107, 13470, 16469, 110, 1543, 51394], "temperature": 0, "avg_logprob": -0.10234992672698666, "compression_ratio": 1.2086720867208671, "no_speech_prob": 1.2707303959080818e-11}, {"id": 5, "seek": 2516, "start": 46.84, "end": 55.0, "text": "再说几个功能聚焦,但场景窄的。Faster Whisper用C-Translate 2加速,速度是原版4倍,内存更低,支持8位量化。", "tokens": [51449, 8623, 8090, 6336, 254, 7549, 32311, 8225, 8171, 248, 4971, 99, 11, 8395, 50255, 50218, 30771, 226, 1546, 1543, 37, 1727, 41132, 610, 9254, 34, 12, 33339, 17593, 568, 9990, 31217, 11, 31217, 13127, 1541, 19683, 42096, 19, 35477, 11, 34742, 39781, 19002, 41377, 11, 35488, 23, 11160, 26748, 23756, 1543, 51857], "temperature": 0, "avg_logprob": -0.10234992672698666, "compression_ratio": 1.2086720867208671, "no_speech_prob": 1.2707303959080818e-11}, {"id": 6, "seek": 5516, "start": 55.419999999999995, "end": 60.5, "text": "行业里用得非常广,但它只是推理引擎,不提供流适输出,你得自己组合其他工具。", "tokens": [50378, 8082, 940, 248, 15759, 9254, 5916, 14392, 3509, 123, 11, 8395, 11284, 36859, 33597, 13876, 41301, 13167, 236, 11, 1960, 20949, 3254, 249, 27854, 2215, 224, 9830, 241, 7781, 11, 2166, 5916, 17645, 10115, 226, 14245, 9572, 5000, 23323, 39806, 1543, 50632], "temperature": 0, "avg_logprob": -0.1063672092982701, "compression_ratio": 1.2404692082111437, "no_speech_prob": 1.3100900572715624e-11}, {"id": 7, "seek": 5516, "start": 60.879999999999995, "end": 66.7, "text": "YT Whisper只做YouTube字幕,场景太窄。Whisper Transparent只做时间戳,没有加速和部署。", "tokens": [50651, 56, 51, 41132, 610, 14003, 10907, 3223, 41173, 22381, 40211, 11, 50255, 50218, 9455, 30771, 226, 1543, 2471, 271, 610, 6531, 38321, 14003, 10907, 44848, 1486, 111, 11, 17944, 9990, 31217, 12565, 13470, 16469, 110, 1543, 50942], "temperature": 0, "avg_logprob": -0.1063672092982701, "compression_ratio": 1.2404692082111437, "no_speech_prob": 1.3100900572715624e-11}, {"id": 8, "seek": 5516, "start": 67.84, "end": 72.92, "text": "实时转录那一批呢?Whisper Mic场景最贴合,麦克风输入实时转录,但只支持英文。", "tokens": [50999, 24726, 15729, 17819, 105, 7391, 243, 4184, 2257, 3416, 117, 6240, 30, 2471, 271, 610, 5818, 50255, 50218, 8661, 18464, 112, 14245, 11, 3994, 99, 24881, 47209, 9830, 241, 14028, 24726, 15729, 17819, 105, 7391, 243, 11, 8395, 14003, 35488, 27869, 17174, 1543, 51253], "temperature": 0, "avg_logprob": -0.1063672092982701, "compression_ratio": 1.2404692082111437, "no_speech_prob": 1.3100900572715624e-11}, {"id": 9, "seek": 5516, "start": 73.38, "end": 79.66, "text": "Whisper Real Time和Whisper Live思路类似,都是边录边识别,但都没有加速优化,中文延迟偏高。", "tokens": [51276, 2471, 271, 610, 8467, 6161, 12565, 2471, 271, 610, 10385, 8870, 24658, 22113, 119, 7384, 120, 11, 22796, 40503, 7391, 243, 40503, 5233, 228, 18453, 11, 8395, 7182, 17944, 9990, 31217, 7384, 246, 23756, 11, 5975, 17174, 17527, 114, 3316, 253, 7437, 237, 12979, 1543, 51590], "temperature": 0, "avg_logprob": -0.1063672092982701, "compression_ratio": 1.2404692082111437, "no_speech_prob": 1.3100900572715624e-11}, {"id": 10, "seek": 7966, "start": 79.66, "end": 85.75999999999999, "text": "Whisper Realization倒是厉害,能识别谁在什么时候说了什么,但串了四五个模型,太重了。", "tokens": [50365, 2471, 271, 610, 8467, 2144, 30878, 1541, 5014, 231, 14694, 11, 8225, 5233, 228, 18453, 33556, 3581, 10440, 29111, 8090, 2289, 10440, 11, 8395, 940, 110, 2289, 19425, 21001, 7549, 41908, 39823, 11, 9455, 12624, 2289, 1543, 50670], "temperature": 0, "avg_logprob": -0.10084469774936108, "compression_ratio": 1.0133928571428572, "no_speech_prob": 1.922861002368048e-11}, {"id": 11, "seek": 7966, "start": 86.61999999999999, "end": 90.25999999999999, "text": "然后是两个重量级选手,Whisper X,综合评价最高。", "tokens": [50713, 26636, 1541, 36257, 7549, 12624, 26748, 16853, 100, 2215, 231, 11389, 11, 2471, 271, 610, 1783, 11, 10115, 120, 14245, 5233, 226, 1550, 115, 8661, 12979, 1543, 50895], "temperature": 0, "avg_logprob": -0.10084469774936108, "compression_ratio": 1.0133928571428572, "no_speech_prob": 1.922861002368048e-11}, {"id": 12, "seek": 7966, "start": 90.75999999999999, "end": 94.52, "text": "Large VR做70倍实时速度,不到8GB显存就能跑。", "tokens": [50920, 43, 289, 432, 13722, 10907, 5867, 35477, 24726, 15729, 31217, 13127, 11, 35728, 23, 8769, 1431, 122, 39781, 3111, 8225, 32585, 1543, 51108], "temperature": 0, "avg_logprob": -0.10084469774936108, "compression_ratio": 1.0133928571428572, "no_speech_prob": 1.922861002368048e-11}, {"id": 13, "seek": 9452, "start": 94.52, "end": 104.92, "text": "关键是支持单词及时间戳,每个字精确对应到音频时间位置,做流逝展示和字幕对齐,这个能力太重要了,短期方案它是最稳的。", "tokens": [50365, 28053, 23049, 106, 1541, 35488, 47446, 5233, 235, 25703, 44848, 1486, 111, 11, 23664, 7549, 22381, 34910, 38114, 106, 8713, 44297, 4511, 18034, 39752, 44848, 11160, 34719, 11, 10907, 27854, 2215, 251, 43491, 25696, 12565, 22381, 40211, 8713, 44628, 238, 11, 15368, 8225, 13486, 9455, 24928, 2289, 11, 5881, 255, 16786, 9249, 28899, 11284, 1541, 8661, 10415, 111, 1546, 1543, 50885], "temperature": 0, "avg_logprob": -0.08734597037820255, "compression_ratio": 1.1797752808988764, "no_speech_prob": 1.763563947654312e-11}, {"id": 14, "seek": 9452, "start": 105.89999999999999, "end": 113.38, "text": "另一个是Distil Whisper,把Large VR蒸留成小模型,速度快六倍,体积小一半,错误率只涨不到1%,性价比太炸了。", "tokens": [50934, 22762, 20182, 1541, 35, 468, 388, 41132, 610, 11, 16075, 43, 289, 432, 13722, 42356, 116, 24456, 11336, 7322, 41908, 39823, 11, 31217, 13127, 10251, 30566, 35477, 11, 29485, 8204, 107, 7322, 2257, 30018, 11, 29900, 5233, 107, 44866, 14003, 35681, 101, 35728, 16, 8923, 21686, 1550, 115, 11706, 9455, 4804, 116, 2289, 1543, 51308], "temperature": 0, "avg_logprob": -0.08734597037820255, "compression_ratio": 1.1797752808988764, "no_speech_prob": 1.763563947654312e-11}, {"id": 15, "seek": 9452, "start": 113.82, "end": 117.39999999999999, "text": "但知名问题是,只发布了英文模型,其他语言得自己蒸留。", "tokens": [51330, 8395, 6498, 15940, 34069, 1541, 11, 14003, 28926, 34688, 2289, 27869, 17174, 41908, 39823, 11, 9572, 5000, 5233, 255, 12009, 5916, 17645, 42356, 116, 24456, 1543, 51509], "temperature": 0, "avg_logprob": -0.08734597037820255, "compression_ratio": 1.1797752808988764, "no_speech_prob": 1.763563947654312e-11}, {"id": 16, "seek": 9452, "start": 117.82, "end": 120.17999999999999, "text": "门槛是有的,但走通了长期收益最大。", "tokens": [51530, 8259, 101, 24880, 249, 1541, 2412, 1546, 11, 8395, 9575, 19550, 2289, 32271, 16786, 18681, 5419, 232, 8661, 3582, 1543, 51648], "temperature": 0, "avg_logprob": -0.08734597037820255, "compression_ratio": 1.1797752808988764, "no_speech_prob": 1.763563947654312e-11}, {"id": 17, "seek": 12018, "start": 120.18, "end": 126.2, "text": "所以方向很清晰,多语言需求选Whisper没毛病,但具体用哪个项目得看你的场景。", "tokens": [50365, 7239, 9249, 24282, 4563, 21784, 5094, 108, 11, 6392, 5233, 255, 12009, 32535, 32718, 2215, 231, 2471, 271, 610, 10062, 39057, 30986, 11, 8395, 39806, 29485, 9254, 17028, 7549, 10178, 117, 11386, 5916, 4200, 18961, 50255, 50218, 1543, 50666], "temperature": 0, "avg_logprob": -0.04834590566919205, "compression_ratio": 1.116751269035533, "no_speech_prob": 1.2509933193516343e-11}, {"id": 18, "seek": 12018, "start": 126.76, "end": 130.5, "text": "短期上Whisper X最稳,长期自蒸留Distil Whisper收益最大。", "tokens": [50694, 5881, 255, 16786, 5708, 2471, 271, 610, 1783, 8661, 10415, 111, 11, 32271, 16786, 9722, 42356, 116, 24456, 35, 468, 388, 41132, 610, 18681, 5419, 232, 8661, 3582, 1543, 50881], "temperature": 0, "avg_logprob": -0.04834590566919205, "compression_ratio": 1.116751269035533, "no_speech_prob": 1.2509933193516343e-11}, {"id": 19, "seek": 12018, "start": 131.0, "end": 133.02, "text": "选项目不是选最好的,是选最合适的。", "tokens": [50906, 2215, 231, 10178, 117, 11386, 7296, 2215, 231, 8661, 20715, 11, 1541, 2215, 231, 8661, 14245, 2215, 224, 1546, 1543, 51007], "temperature": 0, "avg_logprob": -0.04834590566919205, "compression_ratio": 1.116751269035533, "no_speech_prob": 1.2509933193516343e-11}], "language": "zh"}