{"text": "有人把Whisper做到了89ms中位延迟,还顺手把GPU峰值内存砍掉48%。Whisper离线转写很强,但一做实时字幕,就会遇到假设飘移,重算变重,静音误触发。Whisper pipe最聪明的地方不是换模型,而是把实时转写拆成一条可控流水线。它用双缓冲把已确认文本和未确认音频分开,活动音频窗口最多30秒,历史不会越背越重。更关键的是词集时间戳,哪段文字提交了,就按最后一个词的结束时间,把旧音频精准裁掉。提交也不是拍脑袋,简单句走快速通道,模糊句走三次确认,字幕少闪,延迟也压住。结果很硬,P90延迟142ms,平均延迟下降81.1%,WE2只比离线Whisper高约2%。资源更夸张,GPU峰值内存降低48%,平均利用率降低80.9%,150分钟内存零增长。现在用它,别走失效的圆码安装,直接Pipe install Whisper Pipe,然后用Pipe Stream启动。所以Whisper Pipe最值得看的不是单个数字,而是它让实时转写像一个能长期工作的系统。关注置身工坊,继续拆语音AI和开源项目。感谢观看", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 19.32, "text": "有人把Whisper做到了89ms中位延迟,还顺手把GPU峰值内存砍掉48%。Whisper离线转写很强,但一做实时字幕,就会遇到假设飘移,重算变重,静音误触发。Whisper pipe最聪明的地方不是换模型,而是把实时转写拆成一条可控流水线。", "tokens": [50365, 45820, 16075, 2471, 271, 610, 10907, 21381, 21115, 2592, 5975, 11160, 17527, 114, 3316, 253, 11, 14852, 10178, 118, 11389, 16075, 38, 8115, 26555, 108, 40242, 34742, 39781, 23230, 235, 29327, 13318, 4, 1543, 2471, 271, 610, 18614, 119, 16853, 123, 17819, 105, 5676, 247, 4563, 5702, 118, 11, 8395, 2257, 10907, 24726, 15729, 22381, 40211, 11, 3111, 12949, 3330, 229, 4511, 31706, 7422, 122, 11808, 246, 8204, 119, 11, 12624, 19497, 2129, 246, 12624, 11, 5363, 247, 18034, 5233, 107, 6758, 99, 28926, 1543, 2471, 271, 610, 11240, 8661, 8171, 103, 11100, 1546, 30146, 7296, 26075, 95, 41908, 39823, 11, 11070, 1541, 16075, 24726, 15729, 17819, 105, 5676, 247, 6852, 228, 11336, 2257, 48837, 4429, 48707, 27854, 15590, 16853, 123, 1543, 51331], "temperature": 0, "avg_logprob": -0.0793113491751931, "compression_ratio": 1.181547619047619, "no_speech_prob": 8.655473039687056e-12}, {"id": 1, "seek": 0, "start": 20.32, "end": 26.46, "text": "它用双缓冲把已确认文本和未确认音频分开,活动音频窗口最多30秒,历史不会越背越重。", "tokens": [51381, 11284, 9254, 2129, 234, 38109, 241, 5676, 110, 16075, 15701, 38114, 106, 7422, 97, 17174, 8802, 12565, 29954, 38114, 106, 7422, 97, 18034, 39752, 6627, 18937, 11, 25956, 34961, 18034, 39752, 30771, 245, 18144, 8661, 6392, 3446, 20956, 11, 5014, 228, 45399, 47928, 25761, 46329, 25761, 12624, 1543, 51688], "temperature": 0, "avg_logprob": -0.0793113491751931, "compression_ratio": 1.181547619047619, "no_speech_prob": 8.655473039687056e-12}, {"id": 2, "seek": 2646, "start": 26.46, "end": 34.120000000000005, "text": "更关键的是词集时间戳,哪段文字提交了,就按最后一个词的结束时间,把旧音频精准裁掉。", "tokens": [50365, 19002, 28053, 23049, 106, 24620, 5233, 235, 26020, 44848, 1486, 111, 11, 17028, 28427, 17174, 22381, 20949, 28455, 2289, 11, 3111, 26613, 8661, 13547, 20182, 5233, 235, 1546, 45641, 44320, 44848, 11, 16075, 4479, 100, 18034, 39752, 34910, 6336, 228, 8083, 223, 29327, 1543, 50748], "temperature": 0, "avg_logprob": -0.102517713349441, "compression_ratio": 1.111913357400722, "no_speech_prob": 1.5042148082677897e-11}, {"id": 3, "seek": 2646, "start": 35.38, "end": 41.2, "text": "提交也不是拍脑袋,简单句走快速通道,模糊句走三次确认,字幕少闪,延迟也压住。", "tokens": [50811, 20949, 28455, 6404, 7296, 31113, 27067, 239, 19688, 233, 11, 11249, 222, 47446, 34592, 9575, 10251, 31217, 19550, 6025, 11, 41908, 15239, 232, 34592, 9575, 10960, 9487, 38114, 106, 7422, 97, 11, 22381, 40211, 15686, 8259, 103, 11, 17527, 114, 3316, 253, 6404, 5014, 233, 21632, 1543, 51102], "temperature": 0, "avg_logprob": -0.102517713349441, "compression_ratio": 1.111913357400722, "no_speech_prob": 1.5042148082677897e-11}, {"id": 4, "seek": 2646, "start": 42.3, "end": 49.38, "text": "结果很硬,P90延迟142ms,平均延迟下降81.1%,WE2只比离线Whisper高约2%。", "tokens": [51157, 45641, 9319, 4563, 38114, 105, 11, 47, 7771, 17527, 114, 3316, 253, 7271, 17, 2592, 11, 16716, 14872, 229, 17527, 114, 3316, 253, 4438, 47421, 32875, 13, 16, 8923, 37937, 17, 14003, 11706, 18614, 119, 16853, 123, 2471, 271, 610, 12979, 16853, 99, 17, 4, 1543, 51511], "temperature": 0, "avg_logprob": -0.102517713349441, "compression_ratio": 1.111913357400722, "no_speech_prob": 1.5042148082677897e-11}, {"id": 5, "seek": 4938, "start": 49.38, "end": 56.78, "text": "资源更夸张,GPU峰值内存降低48%,平均利用率降低80.9%,150分钟内存零增长。", "tokens": [50365, 5266, 226, 47402, 19002, 1787, 116, 44059, 11, 38, 8115, 26555, 108, 40242, 34742, 39781, 47421, 41377, 13318, 8923, 16716, 14872, 229, 23700, 9254, 44866, 47421, 41377, 4702, 13, 24, 8923, 20120, 6627, 50064, 34742, 39781, 6306, 114, 24228, 252, 32271, 1543, 50735], "temperature": 0, "avg_logprob": -0.10327505111694336, "compression_ratio": 1.134375, "no_speech_prob": 1.3908199245071806e-11}, {"id": 6, "seek": 4938, "start": 57.86, "end": 63.160000000000004, "text": "现在用它,别走失效的圆码安装,直接Pipe install Whisper Pipe,然后用Pipe Stream启动。", "tokens": [50789, 25040, 9254, 11284, 11, 18453, 9575, 28564, 43076, 1546, 2523, 228, 23230, 223, 16206, 8083, 227, 11, 43297, 47, 6527, 3625, 41132, 610, 430, 6527, 11, 26636, 9254, 47, 6527, 24904, 2392, 107, 34961, 1543, 51054], "temperature": 0, "avg_logprob": -0.10327505111694336, "compression_ratio": 1.134375, "no_speech_prob": 1.3908199245071806e-11}, {"id": 7, "seek": 4938, "start": 64.26, "end": 70.04, "text": "所以Whisper Pipe最值得看的不是单个数字,而是它让实时转写像一个能长期工作的系统。", "tokens": [51109, 7239, 2471, 271, 610, 430, 6527, 8661, 40242, 5916, 4200, 1546, 7296, 47446, 7549, 33188, 22381, 11, 11070, 1541, 11284, 33650, 24726, 15729, 17819, 105, 5676, 247, 12760, 20182, 8225, 32271, 16786, 41315, 1546, 25368, 10115, 253, 1543, 51398], "temperature": 0, "avg_logprob": -0.10327505111694336, "compression_ratio": 1.134375, "no_speech_prob": 1.3908199245071806e-11}, {"id": 8, "seek": 4938, "start": 71.14, "end": 74.16, "text": "关注置身工坊,继续拆语音AI和开源项目。", "tokens": [51453, 28053, 26432, 34719, 19847, 23323, 14872, 232, 11, 10115, 100, 10115, 255, 6852, 228, 5233, 255, 18034, 48698, 12565, 18937, 47402, 10178, 117, 11386, 1543, 51604], "temperature": 0, "avg_logprob": -0.10327505111694336, "compression_ratio": 1.134375, "no_speech_prob": 1.3908199245071806e-11}, {"id": 9, "seek": 7416, "start": 74.16, "end": 104.14, "text": "感谢观看", "tokens": [50365, 9709, 11340, 42350, 4200, 51864], "temperature": 0, "avg_logprob": -0.8459521021161761, "compression_ratio": 0.5714285714285714, "no_speech_prob": 3.9292485931596843e-10}], "language": "zh"}