跳转至

全基线公共 Benchmark · 22 点 Pareto 地图

状态:passed。这是 outcome-aware 的探索性整合,不是事前冻结的冠军实验;六个视图在全矩阵派生前固定,没有单一综合分。

22 个 operating points

repr 与完整 wire 分列;worst ΔER 是 Qwen/Fun 的中英与 Parakeet 英文五个 mean 中最差者;state 是已计 learned decoder tensor/state,不代表完整程序/RAM/算力。速度因协议和硬件异构,不进入本表 Pareto。

point repr kbps wire kbps overhead state MiB ViSQOL STOI worst ΔER WavLM speaker fronts
codec2_1200bps 1.198 1.733 44.7% 0.0 3.711 0.630 0.026 0.908 0.711 5/6
codec2_2400bps 2.398 2.935 22.4% 0.0 3.734 0.632 0.026 0.913 0.718 1/6
codec2_3200bps 3.198 3.734 16.8% 0.0 3.740 0.642 0.053 0.932 0.747 1/6
dac_6kbps 6.003 10.320 71.9% 283.0 4.549 0.973 0.011 0.994 0.991 6/6
encodec_12kbps 12.009 12.859 7.1% 88.8 4.311 0.943 0.018 0.983 0.984 3/6
encodec_1_5kbps 1.501 2.346 56.3% 88.8 4.136 0.816 0.130 0.913 0.846 1/6
encodec_24kbps 24.019 24.869 3.5% 88.8 4.331 0.956 0.025 0.987 0.989 2/6
encodec_3kbps 3.002 3.847 28.1% 88.8 4.230 0.874 0.019 0.958 0.930 2/6
encodec_6kbps 6.005 6.850 14.1% 88.8 4.275 0.915 0.021 0.974 0.969 2/6
opus_12kbps 12.038 13.845 15.0% 0.0 4.302 0.961 0.014 0.990 0.984 2/6
opus_24kbps 24.075 25.884 7.5% 0.0 4.266 0.991 0.011 0.997 0.994 6/6
opus_6kbps 6.019 7.822 30.0% 0.0 3.785 0.849 0.015 0.950 0.861 1/6
wavtokenizer_large_unified_480bps 0.481 1.806 275.6% 324.7 4.137 0.835 0.307 0.911 0.861 6/6
moss_v1_q6_750bps 0.754 2.171 188.0% 3387.3 4.205 0.915 0.030 0.981 0.959 5/6
moss_v1_q8_1000bps 1.005 2.424 141.1% 3387.3 4.236 0.930 0.015 0.987 0.972 6/6
moss_v1_q16_2000bps 2.011 3.430 70.6% 3387.3 4.290 0.956 0.004 0.994 0.988 6/6
moss_v1_q32_4000bps 4.022 5.441 35.3% 3387.3 4.326 0.970 0.010 0.996 0.994 6/6
quark_hcodec_1_0_fixed_2000bps 2.005 3.261 62.7% 555.1 4.501 0.928 0.015 0.985 0.976 5/6
quark_hcodec_1_5_threshold_0_4 0.915 2.564 180.1% 2620.0 4.316 0.853 0.050 0.967 0.945 1/6
quark_hcodec_1_5_threshold_0_6 1.451 3.101 113.7% 2620.0 4.403 0.905 0.032 0.982 0.966 4/6
quark_hcodec_1_5_threshold_0_8 1.931 3.582 85.5% 2620.0 4.424 0.919 0.026 0.983 0.969 2/6
bigcodec_official_1040bps 1.041 2.125 104.1% 316.3 4.377 0.920 0.027 0.974 0.962 6/6

六个探索性视图

view frontier dominated edges
representation_quality_all 12 10 54
wire_quality_all 11 11 41
wire_content_en 14 8 13
wire_content_zh 14 8 20
wire_bilingual_content 8 14 42
minimum_sufficient_speech 20 2 2

representation_quality_all

Separate learned representation efficiency from serialized-container overhead.

Frontier:dac_6kbps, opus_24kbps, wavtokenizer_large_unified_480bps, moss_v1_q6_750bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, quark_hcodec_1_0_fixed_2000bps, quark_hcodec_1_5_threshold_0_4, quark_hcodec_1_5_threshold_0_6, quark_hcodec_1_5_threshold_0_8, bigcodec_official_1040bps

wire_quality_all

Compare complete source-blind wire artifacts against automatic quality and speaker proxies.

Frontier:codec2_1200bps, dac_6kbps, opus_24kbps, wavtokenizer_large_unified_480bps, moss_v1_q6_750bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, quark_hcodec_1_0_fixed_2000bps, quark_hcodec_1_5_threshold_0_6, bigcodec_official_1040bps

wire_content_en

English wire/content frontier under all three open evaluator families plus STOI.

Frontier:codec2_1200bps, dac_6kbps, encodec_12kbps, encodec_24kbps, encodec_3kbps, encodec_6kbps, opus_24kbps, wavtokenizer_large_unified_480bps, moss_v1_q6_750bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, quark_hcodec_1_0_fixed_2000bps, bigcodec_official_1040bps

wire_content_zh

Mandarin wire/content frontier under the two supported independent open evaluator families plus STOI.

Frontier:codec2_1200bps, dac_6kbps, encodec_12kbps, opus_12kbps, opus_24kbps, wavtokenizer_large_unified_480bps, moss_v1_q6_750bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, quark_hcodec_1_0_fixed_2000bps, quark_hcodec_1_5_threshold_0_6, quark_hcodec_1_5_threshold_0_8, bigcodec_official_1040bps

wire_bilingual_content

Conservative bilingual content view using the worst supported open-evaluator delta and weaker-language STOI.

Frontier:codec2_1200bps, dac_6kbps, opus_24kbps, wavtokenizer_large_unified_480bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, bigcodec_official_1040bps

minimum_sufficient_speech

Project north-star exploratory view: wire, worst open content, weaker-language intelligibility and speaker proxy, plus charged decoder learned state.

Frontier:codec2_1200bps, codec2_2400bps, codec2_3200bps, dac_6kbps, encodec_12kbps, encodec_1_5kbps, encodec_24kbps, encodec_3kbps, encodec_6kbps, opus_12kbps, opus_24kbps, opus_6kbps, wavtokenizer_large_unified_480bps, moss_v1_q6_750bps, moss_v1_q8_1000bps, moss_v1_q16_2000bps, moss_v1_q32_4000bps, quark_hcodec_1_0_fixed_2000bps, quark_hcodec_1_5_threshold_0_6, bigcodec_official_1040bps

解释边界

  • 现代内容轴统一使用 Qwen3-ASR、Fun-ASR-Nano 与 English-only Parakeet;Whisper/SenseVoice 只保留历史桥接。
  • Parakeet 中文缺失不进入中文视图,也不填 0;负 ΔER 不解释为恢复真实内容。
  • mean dominance 不是统计显著性;各 source aggregate 的 95% CI 保留在 points.jsonl
  • 不同 timing profile 只按 cohort 报告,禁止跨 Mac/GPU、冷/热启动和实现直接排速度名次。
  • 自动指标不替代内容/身份/表达/总体四维人工终审,也不证明 streaming、弱网、实体安全或论文原生复现。

完整 dominance edge 在 dominance_edges.jsonl;Agent Baseline 的选择必须按冻结的 role/feasibility 规则单独裁决,不能按 fronts 次数投票。