{ "text": "支持离线大模型语音识别SOC方案。\n加我微信HAZX17可申请加入数字音频答义群课题组。\n提到的需求,即离线,接入专家模型,支持长短距识别,正是当前端侧AI芯片发展的核心方向。\n这已经不再是简单的关键词唤醒,而是要求芯片能在本地独立运行自动语音识别甚至大语言模型,从而理解并执行复杂的自然语言指令。\n实现这一目标的关键,在于芯片是否集成了强大的神经网络处理单元。\n以下是一些能够满足你需求的SOC芯片,模块级方案,刚性能SOC方案。\n这类芯片通常具备数T OPS的NPU算力,足以在本地运行参数规模,在0.5B到7B之间的大语言模型可现真正的专家模型识别。\n爱心原制AX630C,这是M5Stack Module LLM模块所采用的芯片,集成了3.2TUPS的NPU,并原生支持Transformer模型。\n它配备4GB LPDDR4内存,出厂及预装了QNR,0.5B大语言模型,内置了KWS,ASR,LLM及TTS功能。\n可以直接进行离线语音交互,非常适合开发智能语音助手或智能家居控制中心。\n锐行为2K3576,2K3588,这两款芯片带有6T OPS的NPU,能够支持端侧部署0.5B到3B参数级别的主流大语言模型。\n社区已有基于RK3576NPU实现的全离线智能语音交互系统项目,集成了流逝ASR,DeepSeq大模型推理和TTS门线,实现了语音输入,LLM思考语音输出的完整闭环。\n端到端响应在4米内,耐能Kaneuron KL1140。\n这款芯片宣称是首款能在边缘运行完整Transformer网络的NPU芯片,它专为实时自然语言处理。\n语音界面等应用设计,很新,通过多芯片协同工作,甚至可以运行高达1200亿参数的模型。\n海思HIGH3403\n海思的HIGH3403V100平台提供了一个本地对话智能体方案,在端侧部署了Mini CPM 4V0。\n5B模型实现了ASR,VOM和TTS的完整链路,无需联网即可实现多模态业务识别。\n芯芯威2K1820,2K1828M,2.算力卡。\n这是一款采用M.2接口的AI算力卡。\n搭载了20TOPS的NPU,可独立离线运行3B、7B级的大语言模型。\n金融QN、LMA等主流开源模型。\n为现有设备扩展端侧大模型能力提供了灵活的解决方案,中等算力与低功耗方案。\n这类方案可能在模型规模上有所限制,但依然能够实现比简单KWS复杂得多的语音识别和理解功能。\n瑞玉AmiroC\n瑞玉的AI Voice是一套纯离线的AI解决方案,包含声学信号处理、唤醒、端点检测和识别等模块。\n可以在Amibus Source上搭建复杂的语音语物。\n既可作为纯离线方案,也能实现离在线混合交互。\n支存科技WTM21,这款芯片的特点是超低功耗。\n在实现三百个词的连续识别时,不觉流年赎尽我白发受却。\n千又绕枝叶,醉酒化地面,千水此天又圆着人间。\n我鸟鸟一幅春山空,我老一掉入湖中。\n鸣尘双流所浓,醒来一场春梦。\n我这一生无心踪,我叹事千春千冲。\n谁疯狂的味道是谁在等候。\n二、全航功耗有信用。\n支存WTM211的芯片,在节级功耗上提供了两朝KMOS的试验。\n适合对续航有严苛而成为。\n三、我们考虑集成方式,一无声刺剑。\n不觉流年赎尽我白发就绝。\n千又绕枝叶,醉酒化地面,千又绕枝叶。", "language": "zh", "duration": 288.879438, "segments": [ { "id": 0, "seek": 0, "start": 0.0, "end": 3.2, "text": "支持离线大模型语音识别SOC方案。", "tokens": [ 50365, 35488, 18614, 119, 16853, 123, 3582, 41908, 39823, 5233, 255, 18034, 5233, 228, 18453, 17188, 34, 9249, 28899, 1543, 50525 ], "temperature": 0, "avg_logprob": -0.13990740723662323, "compression_ratio": 1.2546583850931676, "no_speech_prob": 9.395137445800117e-12, "source": "whisper" }, { "id": 1, "seek": 0, "start": 3.2, "end": 8.6, "text": "加我微信HAZX17可申请加入数字音频答义群课题组。", "tokens": [ 50525, 9990, 1654, 39152, 17665, 4983, 57, 55, 7773, 4429, 3526, 111, 27908, 43878, 33188, 22381, 18034, 39752, 28223, 2930, 231, 7248, 97, 5233, 122, 30716, 10115, 226, 1543, 50795 ], "temperature": 0, "avg_logprob": -0.13990740723662323, "compression_ratio": 1.2546583850931676, "no_speech_prob": 9.395137445800117e-12, "source": "whisper" }, { "id": 2, "seek": 0, "start": 8.6, "end": 17.400000000000002, "text": "提到的需求,即离线,接入专家模型,支持长短距识别,正是当前端侧AI芯片发展的核心方向。", "tokens": [ 50795, 20949, 4511, 1546, 32535, 32718, 11, 39127, 18614, 119, 16853, 123, 11, 14468, 14028, 940, 241, 5155, 41908, 39823, 11, 35488, 32271, 5881, 255, 6563, 251, 5233, 228, 18453, 11, 15789, 1541, 16233, 8945, 11957, 107, 3254, 100, 48698, 13778, 107, 16668, 28926, 43491, 1546, 8987, 116, 7945, 9249, 24282, 1543, 51235 ], "temperature": 0, "avg_logprob": -0.13990740723662323, "compression_ratio": 1.2546583850931676, "no_speech_prob": 9.395137445800117e-12, "source": "whisper" }, { "id": 3, "seek": 0, "start": 17.400000000000002, "end": 28.2, "text": "这已经不再是简单的关键词唤醒,而是要求芯片能在本地独立运行自动语音识别甚至大语言模型,从而理解并执行复杂的自然语言指令。", "tokens": [ 51235, 5562, 49161, 1960, 8623, 1541, 11249, 222, 47446, 1546, 28053, 23049, 106, 5233, 235, 17198, 97, 40796, 11, 11070, 1541, 4275, 32718, 13778, 107, 16668, 8225, 3581, 8802, 10928, 18637, 105, 24409, 3316, 238, 8082, 9722, 34961, 5233, 255, 18034, 5233, 228, 18453, 17556, 20844, 3582, 5233, 255, 12009, 41908, 39823, 11, 35630, 11070, 13876, 17278, 3509, 114, 3416, 100, 8082, 1787, 235, 4422, 224, 1546, 9722, 5823, 5233, 255, 12009, 25922, 49061, 1543, 51775 ], "temperature": 0, "avg_logprob": -0.13990740723662323, "compression_ratio": 1.2546583850931676, "no_speech_prob": 9.395137445800117e-12, "source": "whisper" }, { "id": 4, "seek": 2820, "start": 28.2, "end": 34.4, "text": "实现这一目标的关键,在于芯片是否集成了强大的神经网络处理单元。", "tokens": [ 50365, 24726, 20204, 5562, 2257, 11386, 162, 3921, 1546, 28053, 23049, 106, 11, 3581, 37732, 13778, 107, 16668, 1541, 45334, 26020, 11336, 2289, 5702, 118, 39156, 20432, 30276, 16469, 239, 10115, 250, 1787, 226, 13876, 47446, 14812, 1543, 50675 ], "temperature": 0, "avg_logprob": -0.11921985133164593, "compression_ratio": 1.2014652014652014, "no_speech_prob": 1.990555116571091e-11, "source": "whisper" }, { "id": 5, "seek": 2820, "start": 34.4, "end": 40.6, "text": "以下是一些能够满足你需求的SOC芯片,模块级方案,刚性能SOC方案。", "tokens": [ 50675, 3588, 4438, 1541, 38515, 8225, 1787, 253, 15868, 94, 37236, 2166, 32535, 32718, 1546, 17188, 34, 13778, 107, 16668, 11, 41908, 47734, 16853, 100, 9249, 28899, 11, 49160, 21686, 8225, 17188, 34, 9249, 28899, 1543, 50985 ], "temperature": 0, "avg_logprob": -0.11921985133164593, "compression_ratio": 1.2014652014652014, "no_speech_prob": 1.990555116571091e-11, "source": "whisper" }, { "id": 6, "seek": 2820, "start": 40.6, "end": 51.6, "text": "这类芯片通常具备数T OPS的NPU算力,足以在本地运行参数规模,在0.5B到7B之间的大语言模型可现真正的专家模型识别。", "tokens": [ 50985, 5562, 22113, 119, 13778, 107, 16668, 19550, 11279, 39806, 1787, 229, 33188, 51, 422, 6273, 1546, 45, 8115, 19497, 13486, 11, 37236, 3588, 3581, 8802, 10928, 3316, 238, 8082, 2129, 224, 33188, 6758, 226, 41908, 11, 3581, 15, 13, 20, 33, 4511, 22, 33, 9574, 31685, 1546, 3582, 5233, 255, 12009, 41908, 39823, 4429, 20204, 6303, 15789, 1546, 940, 241, 5155, 41908, 39823, 5233, 228, 18453, 1543, 51535 ], "temperature": 0, "avg_logprob": -0.11921985133164593, "compression_ratio": 1.2014652014652014, "no_speech_prob": 1.990555116571091e-11, "source": "whisper" }, { "id": 7, "seek": 5160, "start": 51.6, "end": 63.34, "text": "爱心原制AX630C,这是M5Stack Module LLM模块所采用的芯片,集成了3.2TUPS的NPU,并原生支持Transformer模型。", "tokens": [ 50365, 27324, 7945, 19683, 25491, 32, 55, 21, 3446, 34, 11, 27455, 44, 20, 4520, 501, 48251, 441, 43, 44, 41908, 47734, 5966, 5873, 229, 9254, 1546, 13778, 107, 16668, 11, 26020, 11336, 2289, 18, 13, 17, 51, 52, 6273, 1546, 45, 8115, 11, 3509, 114, 19683, 8244, 35488, 33339, 837, 260, 41908, 39823, 1543, 50952 ], "temperature": 0, "avg_logprob": -0.17053148664277176, "compression_ratio": 0.9197080291970803, "no_speech_prob": 1.8641569191069074e-11, "source": "whisper" }, { "id": 8, "seek": 6334, "start": 63.34, "end": 75.98, "text": "它配备4GB LPDDR4内存,出厂及预装了QNR,0.5B大语言模型,内置了KWS,ASR,LLM及TTS功能。", "tokens": [ 50365, 11284, 38846, 1787, 229, 19, 8769, 441, 17349, 9301, 19, 34742, 39781, 11, 7781, 5014, 224, 25703, 12501, 226, 8083, 227, 2289, 48, 45, 49, 11, 15, 13, 20, 33, 3582, 5233, 255, 12009, 41908, 39823, 11, 34742, 34719, 2289, 42, 12508, 11, 3160, 49, 11, 43, 43, 44, 25703, 51, 7327, 32311, 8225, 1543, 50997 ], "temperature": 0, "avg_logprob": -0.15112298221911413, "compression_ratio": 0.9035087719298246, "no_speech_prob": 2.0270500555863435e-11, "source": "whisper" }, { "id": 9, "seek": 7598, "start": 75.98, "end": 82.5, "text": "可以直接进行离线语音交互,非常适合开发智能语音助手或智能家居控制中心。", "tokens": [ 50365, 6723, 43297, 36700, 8082, 18614, 119, 16853, 123, 5233, 255, 18034, 28455, 1369, 240, 11, 14392, 2215, 224, 14245, 18937, 28926, 5094, 118, 8225, 5233, 255, 18034, 37618, 11389, 19780, 5094, 118, 8225, 5155, 9636, 227, 48707, 25491, 5975, 7945, 1543, 50691 ], "temperature": 0, "avg_logprob": -0.09225207567214966, "compression_ratio": 1.0357142857142858, "no_speech_prob": 1.6682917200472325e-11, "source": "whisper" }, { "id": 10, "seek": 7598, "start": 83.32000000000001, "end": 94.68, "text": "锐行为2K3576,2K3588,这两款芯片带有6T OPS的NPU,能够支持端侧部署0.5B到3B参数级别的主流大语言模型。", "tokens": [ 50732, 23049, 238, 8082, 13992, 17, 42, 8794, 25026, 11, 17, 42, 8794, 16919, 11, 5562, 36257, 48798, 13778, 107, 16668, 4845, 99, 2412, 21, 51, 422, 6273, 1546, 45, 8115, 11, 8225, 1787, 253, 35488, 11957, 107, 3254, 100, 13470, 16469, 110, 15, 13, 20, 33, 4511, 18, 33, 2129, 224, 33188, 16853, 100, 18453, 1546, 13557, 27854, 3582, 5233, 255, 12009, 41908, 39823, 1543, 51300 ], "temperature": 0, "avg_logprob": -0.09225207567214966, "compression_ratio": 1.0357142857142858, "no_speech_prob": 1.6682917200472325e-11, "source": "whisper" }, { "id": 11, "seek": 9468, "start": 94.68, "end": 110.80000000000001, "text": "社区已有基于RK3576NPU实现的全离线智能语音交互系统项目,集成了流逝ASR,DeepSeq大模型推理和TTS门线,实现了语音输入,LLM思考语音输出的完整闭环。", "tokens": [ 50365, 27658, 9937, 118, 15701, 2412, 26008, 37732, 49, 42, 8794, 25026, 45, 8115, 24726, 20204, 1546, 11319, 18614, 119, 16853, 123, 5094, 118, 8225, 5233, 255, 18034, 28455, 1369, 240, 25368, 10115, 253, 10178, 117, 11386, 11, 26020, 11336, 2289, 27854, 2215, 251, 3160, 49, 11, 11089, 595, 10637, 80, 3582, 41908, 39823, 33597, 13876, 12565, 51, 7327, 8259, 101, 16853, 123, 11, 24726, 20204, 2289, 5233, 255, 18034, 9830, 241, 14028, 11, 24010, 44, 8870, 26504, 5233, 255, 18034, 9830, 241, 7781, 1546, 14128, 27662, 8259, 255, 8051, 107, 1543, 51171 ], "temperature": 0, "avg_logprob": -0.1417200716933763, "compression_ratio": 1.1222570532915361, "no_speech_prob": 1.97893437747787e-11, "source": "whisper" }, { "id": 12, "seek": 9468, "start": 111.26, "end": 115.54, "text": "端到端响应在4米内,耐能Kaneuron KL1140。", "tokens": [ 51194, 11957, 107, 4511, 11957, 107, 3858, 235, 44297, 3581, 19, 27742, 34742, 11, 4450, 238, 8225, 42, 1929, 374, 266, 47991, 5348, 5254, 1543, 51408 ], "temperature": 0, "avg_logprob": -0.1417200716933763, "compression_ratio": 1.1222570532915361, "no_speech_prob": 1.97893437747787e-11, "source": "whisper" }, { "id": 13, "seek": 9468, "start": 115.98, "end": 124.4, "text": "这款芯片宣称是首款能在边缘运行完整Transformer网络的NPU芯片,它专为实时自然语言处理。", "tokens": [ 51430, 5562, 48798, 13778, 107, 16668, 2415, 96, 8204, 108, 1541, 25444, 48798, 8225, 3581, 40503, 38109, 246, 3316, 238, 8082, 14128, 27662, 33339, 837, 260, 16469, 239, 10115, 250, 1546, 45, 8115, 13778, 107, 16668, 11, 11284, 940, 241, 13992, 24726, 15729, 9722, 5823, 5233, 255, 12009, 1787, 226, 13876, 1543, 51851 ], "temperature": 0, "avg_logprob": -0.1417200716933763, "compression_ratio": 1.1222570532915361, "no_speech_prob": 1.97893437747787e-11, "source": "whisper" }, { "id": 14, "seek": 12468, "start": 124.68, "end": 131.70000000000002, "text": "语音界面等应用设计,很新,通过多芯片协同工作,甚至可以运行高达1200亿参数的模型。", "tokens": [ 50365, 5233, 255, 18034, 21185, 8833, 10187, 44297, 9254, 7422, 122, 7422, 94, 11, 4563, 12560, 11, 19550, 16866, 6392, 13778, 107, 16668, 5322, 237, 13089, 41315, 11, 17556, 20844, 6723, 3316, 238, 8082, 12979, 9830, 122, 4762, 628, 1369, 123, 2129, 224, 33188, 1546, 41908, 39823, 1543, 50716 ], "temperature": 0, "avg_logprob": -0.14340223584856307, "compression_ratio": 1.1099656357388317, "no_speech_prob": 1.6665328839149396e-11, "source": "whisper" }, { "id": 15, "seek": 12468, "start": 132.36, "end": 133.70000000000002, "text": "海思HIGH3403", "tokens": [ 50749, 20078, 8870, 39, 15483, 18, 5254, 18, 50816 ], "temperature": 0, "avg_logprob": -0.14340223584856307, "compression_ratio": 1.1099656357388317, "no_speech_prob": 1.6665328839149396e-11, "source": "whisper" }, { "id": 16, "seek": 12468, "start": 133.70000000000002, "end": 141.8, "text": "海思的HIGH3403V100平台提供了一个本地对话智能体方案,在端侧部署了Mini CPM 4V0。", "tokens": [ 50816, 20078, 8870, 1546, 39, 15483, 18, 5254, 18, 53, 6879, 16716, 15433, 20949, 3254, 249, 2289, 20182, 8802, 10928, 8713, 21596, 5094, 118, 8225, 29485, 9249, 28899, 11, 3581, 11957, 107, 3254, 100, 13470, 16469, 110, 2289, 44, 3812, 383, 18819, 1017, 53, 15, 1543, 51221 ], "temperature": 0, "avg_logprob": -0.14340223584856307, "compression_ratio": 1.1099656357388317, "no_speech_prob": 1.6665328839149396e-11, "source": "whisper" }, { "id": 17, "seek": 12468, "start": 142.02, "end": 149.8, "text": "5B模型实现了ASR,VOM和TTS的完整链路,无需联网即可实现多模态业务识别。", "tokens": [ 51232, 20, 33, 41908, 39823, 24726, 20204, 2289, 3160, 49, 11, 53, 5251, 12565, 51, 7327, 1546, 14128, 27662, 165, 241, 122, 24658, 11, 31364, 32535, 8171, 242, 16469, 239, 39127, 4429, 24726, 20204, 6392, 41908, 3757, 223, 940, 248, 5087, 94, 5233, 228, 18453, 1543, 51621 ], "temperature": 0, "avg_logprob": -0.14340223584856307, "compression_ratio": 1.1099656357388317, "no_speech_prob": 1.6665328839149396e-11, "source": "whisper" }, { "id": 18, "seek": 14980, "start": 149.8, "end": 154.5, "text": "芯芯威2K1820,2K1828M,2.算力卡。", "tokens": [ 50365, 13778, 107, 13778, 107, 19976, 223, 17, 42, 6494, 2009, 11, 17, 42, 6494, 11205, 44, 11, 17, 13, 19497, 13486, 32681, 1543, 50600 ], "temperature": 0, "avg_logprob": -0.19746742692104605, "compression_ratio": 0.9936305732484076, "no_speech_prob": 1.7647066100079378e-11, "source": "whisper" }, { "id": 19, "seek": 14980, "start": 155.24, "end": 158.0, "text": "这是一款采用M.2接口的AI算力卡。", "tokens": [ 50637, 27455, 2257, 48798, 5873, 229, 9254, 44, 13, 17, 14468, 18144, 1546, 48698, 19497, 13486, 32681, 1543, 50775 ], "temperature": 0, "avg_logprob": -0.19746742692104605, "compression_ratio": 0.9936305732484076, "no_speech_prob": 1.7647066100079378e-11, "source": "whisper" }, { "id": 20, "seek": 14980, "start": 158.58, "end": 163.84, "text": "搭载了20TOPS的NPU,可独立离线运行3B、7B级的大语言模型。", "tokens": [ 50804, 22220, 255, 17819, 121, 2289, 2009, 15427, 6273, 1546, 45, 8115, 11, 4429, 18637, 105, 24409, 18614, 119, 16853, 123, 3316, 238, 8082, 18, 33, 1231, 22, 33, 16853, 100, 1546, 3582, 5233, 255, 12009, 41908, 39823, 1543, 51067 ], "temperature": 0, "avg_logprob": -0.19746742692104605, "compression_ratio": 0.9936305732484076, "no_speech_prob": 1.7647066100079378e-11, "source": "whisper" }, { "id": 21, "seek": 16384, "start": 163.84, "end": 167.08, "text": "金融QN、LMA等主流开源模型。", "tokens": [ 50365, 19117, 43772, 235, 48, 45, 1231, 43, 9998, 10187, 13557, 27854, 18937, 47402, 41908, 39823, 1543, 50527 ], "temperature": 0, "avg_logprob": -0.14927778822002988, "compression_ratio": 1.1365461847389557, "no_speech_prob": 1.548133496398485e-11, "source": "whisper" }, { "id": 22, "seek": 16384, "start": 167.66, "end": 174.2, "text": "为现有设备扩展端侧大模型能力提供了灵活的解决方案,中等算力与低功耗方案。", "tokens": [ 50556, 13992, 20204, 2412, 7422, 122, 1787, 229, 3416, 102, 43491, 11957, 107, 3254, 100, 3582, 41908, 39823, 8225, 13486, 20949, 3254, 249, 2289, 13290, 113, 25956, 1546, 17278, 5676, 111, 9249, 28899, 11, 5975, 10187, 19497, 13486, 940, 236, 41377, 32311, 4450, 245, 9249, 28899, 1543, 50883 ], "temperature": 0, "avg_logprob": -0.14927778822002988, "compression_ratio": 1.1365461847389557, "no_speech_prob": 1.548133496398485e-11, "source": "whisper" }, { "id": 23, "seek": 16384, "start": 174.72, "end": 182.38, "text": "这类方案可能在模型规模上有所限制,但依然能够实现比简单KWS复杂得多的语音识别和理解功能。", "tokens": [ 50909, 5562, 22113, 119, 9249, 28899, 16657, 3581, 41908, 39823, 6758, 226, 41908, 5708, 2412, 5966, 43446, 25491, 11, 8395, 3254, 251, 5823, 8225, 1787, 253, 24726, 20204, 11706, 11249, 222, 47446, 42, 12508, 1787, 235, 4422, 224, 5916, 6392, 1546, 5233, 255, 18034, 5233, 228, 18453, 12565, 13876, 17278, 32311, 8225, 1543, 51292 ], "temperature": 0, "avg_logprob": -0.14927778822002988, "compression_ratio": 1.1365461847389557, "no_speech_prob": 1.548133496398485e-11, "source": "whisper" }, { "id": 24, "seek": 16384, "start": 183.36, "end": 184.52, "text": "瑞玉AmiroC", "tokens": [ 51341, 37955, 252, 8051, 231, 32, 3057, 340, 34, 51399 ], "temperature": 0, "avg_logprob": -0.14927778822002988, "compression_ratio": 1.1365461847389557, "no_speech_prob": 1.548133496398485e-11, "source": "whisper" }, { "id": 25, "seek": 18452, "start": 184.52, "end": 193.3, "text": "瑞玉的AI Voice是一套纯离线的AI解决方案,包含声学信号处理、唤醒、端点检测和识别等模块。", "tokens": [ 50365, 37955, 252, 8051, 231, 1546, 48698, 15229, 1541, 2257, 1881, 245, 16853, 107, 18614, 119, 16853, 123, 1546, 48698, 17278, 5676, 111, 9249, 28899, 11, 23305, 2392, 104, 32045, 29618, 17665, 26987, 1787, 226, 13876, 1231, 17198, 97, 40796, 1231, 11957, 107, 12579, 16407, 222, 11038, 233, 12565, 5233, 228, 18453, 10187, 41908, 47734, 1543, 50804 ], "temperature": 0, "avg_logprob": -0.1999029530038079, "compression_ratio": 1.099630996309963, "no_speech_prob": 1.8528203277190514e-11, "source": "whisper" }, { "id": 26, "seek": 18452, "start": 193.9, "end": 197.0, "text": "可以在Amibus Source上搭建复杂的语音语物。", "tokens": [ 50834, 6723, 3581, 16833, 897, 301, 29629, 5708, 22220, 255, 34157, 1787, 235, 4422, 224, 1546, 5233, 255, 18034, 5233, 255, 23516, 1543, 50989 ], "temperature": 0, "avg_logprob": -0.1999029530038079, "compression_ratio": 1.099630996309963, "no_speech_prob": 1.8528203277190514e-11, "source": "whisper" }, { "id": 27, "seek": 18452, "start": 197.58, "end": 201.48000000000002, "text": "既可作为纯离线方案,也能实现离在线混合交互。", "tokens": [ 51018, 4479, 95, 4429, 11914, 13992, 16853, 107, 18614, 119, 16853, 123, 9249, 28899, 11, 6404, 8225, 24726, 20204, 18614, 119, 3581, 16853, 123, 48640, 14245, 28455, 1369, 240, 1543, 51213 ], "temperature": 0, "avg_logprob": -0.1999029530038079, "compression_ratio": 1.099630996309963, "no_speech_prob": 1.8528203277190514e-11, "source": "whisper" }, { "id": 28, "seek": 18452, "start": 202.48000000000002, "end": 207.52, "text": "支存科技WTM21,这款芯片的特点是超低功耗。", "tokens": [ 51263, 24400, 39781, 42091, 32502, 54, 42023, 4436, 11, 5562, 48798, 13778, 107, 16668, 1546, 17682, 12579, 1541, 19869, 41377, 32311, 4450, 245, 1543, 51515 ], "temperature": 0, "avg_logprob": -0.1999029530038079, "compression_ratio": 1.099630996309963, "no_speech_prob": 1.8528203277190514e-11, "source": "whisper" }, { "id": 29, "seek": 20752, "start": 207.52, "end": 215.62, "text": "在实现三百个词的连续识别时,不觉流年赎尽我白发受却。", "tokens": [ 50365, 3581, 24726, 20204, 10960, 31906, 7549, 5233, 235, 1546, 3316, 252, 10115, 255, 5233, 228, 18453, 15729, 11, 1960, 24447, 27854, 5157, 5266, 236, 1530, 121, 1654, 13558, 28926, 23151, 5322, 112, 1543, 50770 ], "temperature": 0, "avg_logprob": -0.4681876500447591, "compression_ratio": 1.0, "no_speech_prob": 3.335549891847478e-11, "source": "whisper" }, { "id": 30, "seek": 20752, "start": 216.98000000000002, "end": 228.16000000000003, "text": "千又绕枝叶,醉酒化地面,千水此天又圆着人间。", "tokens": [ 50838, 20787, 17047, 10115, 243, 7360, 251, 2129, 114, 11, 21038, 231, 37486, 23756, 10928, 8833, 11, 20787, 15590, 17947, 6135, 17047, 2523, 228, 20708, 4035, 31685, 1543, 51397 ], "temperature": 0, "avg_logprob": -0.4681876500447591, "compression_ratio": 1.0, "no_speech_prob": 3.335549891847478e-11, "source": "whisper" }, { "id": 31, "seek": 22816, "start": 228.16, "end": 237.12, "text": "我鸟鸟一幅春山空,我老一掉入湖中。", "tokens": [ 50365, 1654, 165, 116, 253, 165, 116, 253, 2257, 3509, 227, 46953, 24113, 23322, 11, 1654, 10439, 2257, 29327, 14028, 33744, 244, 5975, 1543, 50813 ], "temperature": 0, "avg_logprob": -0.5545520237513951, "compression_ratio": 1.0384615384615385, "no_speech_prob": 4.65362644253009e-11, "source": "whisper" }, { "id": 32, "seek": 22816, "start": 238.12, "end": 244.3, "text": "鸣尘双流所浓,醒来一场春梦。", "tokens": [ 50863, 165, 116, 96, 1530, 246, 2129, 234, 27854, 5966, 11038, 241, 11, 40796, 6912, 2257, 50255, 46953, 19017, 99, 1543, 51172 ], "temperature": 0, "avg_logprob": -0.5545520237513951, "compression_ratio": 1.0384615384615385, "no_speech_prob": 4.65362644253009e-11, "source": "whisper" }, { "id": 33, "seek": 22816, "start": 244.3, "end": 251.42, "text": "我这一生无心踪,我叹事千春千冲。", "tokens": [ 51172, 1654, 5562, 2257, 8244, 31364, 7945, 41166, 103, 11, 1654, 2129, 117, 6973, 20787, 46953, 20787, 5676, 110, 1543, 51528 ], "temperature": 0, "avg_logprob": -0.5545520237513951, "compression_ratio": 1.0384615384615385, "no_speech_prob": 4.65362644253009e-11, "source": "whisper" }, { "id": 34, "seek": 25142, "start": 251.42, "end": 258.41999999999996, "text": "谁疯狂的味道是谁在等候。", "tokens": [ 50365, 33556, 18140, 107, 18637, 224, 1546, 17268, 6025, 1541, 33556, 3581, 10187, 11238, 1543, 50715 ], "temperature": 0, "avg_logprob": -0.6353443463643392, "compression_ratio": 0.7659574468085106, "no_speech_prob": 2.2368334615108232e-11, "source": "whisper" }, { "id": 35, "seek": 25842, "start": 258.42, "end": 261.64000000000004, "text": "二、全航功耗有信用。", "tokens": [ 50365, 11217, 1231, 11319, 10256, 103, 32311, 4450, 245, 2412, 17665, 9254, 1543, 50526 ], "temperature": 0, "avg_logprob": -0.551652610848803, "compression_ratio": 1.0776699029126213, "no_speech_prob": 2.320160036317631e-11, "source": "whisper" }, { "id": 36, "seek": 25842, "start": 262.32, "end": 268.38, "text": "支存WTM211的芯片,在节级功耗上提供了两朝KMOS的试验。", "tokens": [ 50560, 24400, 39781, 54, 42023, 4436, 16, 1546, 13778, 107, 16668, 11, 3581, 45161, 16853, 100, 32311, 4450, 245, 5708, 20949, 3254, 249, 2289, 36257, 46610, 42, 44, 4367, 1546, 5233, 243, 49657, 234, 1543, 50863 ], "temperature": 0, "avg_logprob": -0.551652610848803, "compression_ratio": 1.0776699029126213, "no_speech_prob": 2.320160036317631e-11, "source": "whisper" }, { "id": 37, "seek": 25842, "start": 268.78000000000003, "end": 270.8, "text": "适合对续航有严苛而成为。", "tokens": [ 50883, 2215, 224, 14245, 8713, 10115, 255, 10256, 103, 2412, 940, 98, 13736, 249, 11070, 11336, 13992, 1543, 50984 ], "temperature": 0, "avg_logprob": -0.551652610848803, "compression_ratio": 1.0776699029126213, "no_speech_prob": 2.320160036317631e-11, "source": "whisper" }, { "id": 38, "seek": 25842, "start": 270.8, "end": 275.72, "text": "三、我们考虑集成方式,一无声刺剑。", "tokens": [ 50984, 10960, 1231, 15003, 26504, 12026, 239, 26020, 11336, 9249, 27584, 11, 2257, 31364, 32045, 2437, 118, 5935, 239, 1543, 51230 ], "temperature": 0, "avg_logprob": -0.551652610848803, "compression_ratio": 1.0776699029126213, "no_speech_prob": 2.320160036317631e-11, "source": "whisper" }, { "id": 39, "seek": 25842, "start": 275.72, "end": 282.98, "text": "不觉流年赎尽我白发就绝。", "tokens": [ 51230, 1960, 24447, 27854, 5157, 5266, 236, 1530, 121, 1654, 13558, 28926, 3111, 10115, 251, 1543, 51593 ], "temperature": 0, "avg_logprob": -0.551652610848803, "compression_ratio": 1.0776699029126213, "no_speech_prob": 2.320160036317631e-11, "source": "whisper" }, { "id": 40, "seek": 28298, "start": 282.98, "end": 288.86, "text": "千又绕枝叶,醉酒化地面,千又绕枝叶。", "tokens": [ 50365, 20787, 17047, 10115, 243, 7360, 251, 2129, 114, 11, 21038, 231, 37486, 23756, 10928, 8833, 11, 20787, 17047, 10115, 243, 7360, 251, 2129, 114, 1543, 50659 ], "temperature": 0, "avg_logprob": -0.3841104507446289, "compression_ratio": 1.0204081632653061, "no_speech_prob": 3.3086478001820296e-11, "source": "whisper" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 288.879438, "reference_path": null, "reference_kind": null, "reference_language": null, "selected_source": "whisper_segmented", "decision_reason": "source_unavailable_or_untrusted_whisper_passed", "whisper_rejected": false, "source_profile": { "segment_count": 0, "duration_coverage": 0.0, "usable_full": false, "usable_partial": false, "hallucination_region_count": 0, "hallucination_regions": [], "language": { "requested": "zh", "matched": false } }, "whisper_profile": { "segment_count": 41, "multilingual": false, "first_start": 0.0, "final_end": 288.86, "duration_coverage": 0.999933, "maximum_gap_seconds": 1.36, "inverted_count": 0, "language": { "requested": "zh", "matched": true, "latin": 226, "han": 895, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "comparison": { "window_seconds": 30.0, "compared_window_count": 0, "median_score": 0.0, "divergent_window_count": 0, "divergent_ratio": 0.0, "divergent_windows": [] }, "timestamp_alignment": { "applied": false, "estimated_offset_seconds": 0.0, "matched_segment_count": 0, "match_scores": [] }, "caption_timing_repair": null, "filled_source_profile": null, "whisper_duration_guard": { "media_duration": 288.879, "rejected_segment_count": 0, "rejected_segments": [] } } }