.gitattributes CHANGED
@@ -1,36 +1,7 @@
1
- *.7z filter=lfs diff=lfs merge=lfs -text
2
- *.arrow filter=lfs diff=lfs merge=lfs -text
3
  *.bin filter=lfs diff=lfs merge=lfs -text
4
- *.bz2 filter=lfs diff=lfs merge=lfs -text
5
- *.ckpt filter=lfs diff=lfs merge=lfs -text
6
- *.ftz filter=lfs diff=lfs merge=lfs -text
7
- *.gz filter=lfs diff=lfs merge=lfs -text
8
- *.h5 filter=lfs diff=lfs merge=lfs -text
9
- *.joblib filter=lfs diff=lfs merge=lfs -text
10
- *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
- *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
- *.model filter=lfs diff=lfs merge=lfs -text
13
- *.msgpack filter=lfs diff=lfs merge=lfs -text
14
- *.npy filter=lfs diff=lfs merge=lfs -text
15
- *.npz filter=lfs diff=lfs merge=lfs -text
16
- *.onnx filter=lfs diff=lfs merge=lfs -text
17
- *.ot filter=lfs diff=lfs merge=lfs -text
18
  *.parquet filter=lfs diff=lfs merge=lfs -text
19
- *.pb filter=lfs diff=lfs merge=lfs -text
20
- *.pickle filter=lfs diff=lfs merge=lfs -text
21
- *.pkl filter=lfs diff=lfs merge=lfs -text
22
- *.pt filter=lfs diff=lfs merge=lfs -text
23
- *.pth filter=lfs diff=lfs merge=lfs -text
24
- *.rar filter=lfs diff=lfs merge=lfs -text
25
- *.safetensors filter=lfs diff=lfs merge=lfs -text
26
- saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
- *.tar.* filter=lfs diff=lfs merge=lfs -text
28
- *.tar filter=lfs diff=lfs merge=lfs -text
29
- *.tflite filter=lfs diff=lfs merge=lfs -text
30
- *.tgz filter=lfs diff=lfs merge=lfs -text
31
- *.wasm filter=lfs diff=lfs merge=lfs -text
32
- *.xz filter=lfs diff=lfs merge=lfs -text
33
- *.zip filter=lfs diff=lfs merge=lfs -text
34
- *.zst filter=lfs diff=lfs merge=lfs -text
35
- *tfevents* filter=lfs diff=lfs merge=lfs -text
36
- tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
1
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
 
2
  *.bin filter=lfs diff=lfs merge=lfs -text
3
+ *.wav filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  *.parquet filter=lfs diff=lfs merge=lfs -text
5
+ image/01_metric_comparison.png filter=lfs diff=lfs merge=lfs -text
6
+ image/02_latency_comparison.png filter=lfs diff=lfs merge=lfs -text
7
+ image/效果对比.png filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
EVALUATION.md ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Turn-Detection-Jev:训练与评估详情
2
+
3
+ [返回模型主页](README.md)
4
+
5
+ 开源发布名为 **Turn-Detection-Jev**。
6
+
7
+ ## 训练数据与配置
8
+
9
+ 该模型 共使用 **156,925 条**处理后的样本,来源为 Easy-Turn、Smart-Turn 和 MUSAN noise([OpenSLR 17](https://www.openslr.org/17/) 的 noise 子集)。
10
+
11
+ | 来源 | 纳入样本数 |
12
+ | --- | ---: |
13
+ | Easy-Turn | 115,995 |
14
+ | Smart-Turn | 40,000 |
15
+ | MUSAN noise | 930 |
16
+ | **合计** | **156,925** |
17
+
18
+ | 最终标签 | train | validation | calibration | test |
19
+ | --- | ---: | ---: | ---: | ---: |
20
+ | `listen_complete` | 45,000 | 2,500 | 1,250 | 1,250 |
21
+ | `listen_incomplete` | 22,500 | 1,250 | 625 | 625 |
22
+ | `listen_background` | 419 | 24 | 12 | 12 |
23
+ | `speak_interrupt` | 27,000 | 1,500 | 750 | 750 |
24
+ | `speak_backchannel` | 23,395 | 1,300 | 650 | 650 |
25
+ | `speak_incomplete` | 22,500 | 1,250 | 625 | 625 |
26
+ | `speak_background` | 418 | 23 | 11 | 11 |
27
+ | **合计** | **141,232** | **7,847** | **3,923** | **3,923** |
28
+
29
+ 训练过程中将五种语义按照助手实际状态拆分为七小类。
30
+
31
+ | 训练参数 | 值 |
32
+ | --- | --- |
33
+ | LoRA rank / alpha / dropout | 32 / 64 / 0.05 |
34
+ | 学习率 / 训练预算 | 5e-5 / 3 epochs |
35
+ | 单卡 train / eval batch size | 4 / 4 |
36
+ | 梯度累积 / 有效训练 batch | 64 / 256(本次单卡训练) |
37
+ | 精度 / 最大输入 token 数 | bfloat16 / 4096 |
38
+ | 损失 | `audio_jev_ce_brier`,Brier 权重 1.0 |
39
+ | warmup ratio / weight decay | 0.03 / 0 |
40
+ | 随机种子 | 0 |
41
+ | 检查点选择 | 每 100 步验证,以最低验证 NLL 选择 best |
42
+
43
+ ## 评估结果
44
+
45
+ ### 公开测试集:完整评估范围
46
+
47
+ 结果来自 2026-11-11 完整原生 Transformers 运行 `merged-20261111_051726-1672084`,prompt 为 `audio-jev-labels-v4`、评分 head 为 `lm_head`、温度 T=1,未启用 calibration。两套报告均为 `is_full_run=true`。该运行替换主页中的历史 ms-swift 成绩,历史原始记录仍保留在代码仓库 `results/reference/`。
48
+
49
+ 代码仓库 `docs/public-benchmark.json` 保存完整/共同样本的精确指标、预测文件 SHA256 和来源路径;路径均为仓库相对路径。
50
+
51
+ | 数据集 | N | 正确数 | Accuracy | 状态标签 Macro-F1 | 语义 Macro-F1 |
52
+ | --- | ---: | ---: | ---: | ---: | ---: |
53
+ | Easy-Turn | 800 | 789 | 98.625% | 98.83% | 99.00% |
54
+ | Smart-Test-4000 | 4,000 | 3,871 | 96.775% | 96.78% | 96.81% |
55
+
56
+ 公开集宏平均只覆盖有真值样本的类别。语义映射将两种 incomplete 合并为 `incomplete`,两种 background 合并为 `invalid`,`speak_interrupt` 对应 Easy-Turn 原始 `wait`。零支持类别不代表已经获得完整评估。
57
+
58
+ | 状态标签 | Easy-Turn 正确 / N | Recall | Smart-Test 正确 / N | Recall |
59
+ | --- | ---: | ---: | ---: | ---: |
60
+ | `listen_complete` | 296 / 300 | 98.67% | 1,958 / 2,000 | 97.90% |
61
+ | `listen_incomplete` | 144 / 150 | 96.00% | 914 / 1,000 | 91.40% |
62
+ | `listen_background` | 0 / 0 | 未评估 | 0 / 0 | 未评估 |
63
+ | `speak_interrupt` | 100 / 100 | 100.00% | 0 / 0 | 未评估 |
64
+ | `speak_backchannel` | 100 / 100 | 100.00% | 0 / 0 | 未评估 |
65
+ | `speak_incomplete` | 149 / 150 | 99.33% | 999 / 1,000 | 99.90% |
66
+ | `speak_background` | 0 / 0 | 未评估 | 0 / 0 | 未评估 |
67
+
68
+ Easy-Turn 的 listening / speaking 准确率分别为 **97.78%(440/450)**、**99.71%(349/350)**;Smart-Test 分别为 **95.73%(2872/3000)**、**99.90%(999/1000)**。
69
+
70
+ ### 共同样本上的模型参照
71
+
72
+ Easy-Turn 下表只取 complete/incomplete 的 **600 条**共同样本;Turn-Detection-Jev 为 **589/600**,与全量 800 条的 789/800 是不同范围。Smart-Test 使用固定 4,000 条子集。Turn-Detection-Jev 使用本次结果,其他模型保留历史结果。
73
+
74
+ | 数据集 | 模型 | N | 准确率 | 精准率(宏平均) | 召回率(宏平均) | F1(宏平均) |
75
+ | --- | --- | --- | --- | --- | --- | --- |
76
+ | easyturn | TurnSense 1.1 GPU | 600 | 96.17% | 96.19% | 96.17% | 96.17% |
77
+ | easyturn | Smart-Turn 3.2 CPU | 600 | 76.00% | 78.21% | 76.00% | 75.52% |
78
+ | easyturn | Smart-Turn 3.2 GPU | 600 | 77.17% | 78.26% | 77.17% | 76.94% |
79
+ | easyturn | Qwen3-Omni-30B-A3B-Instruct | 600 | 92.33% | 93.40% | 92.33% | 92.53% |
80
+ | easyturn | Turn-Detection-Jev | 600 | 98.17% | 98.33% | 98.17% | 98.25% |
81
+ | smart-test | TurnSense 1.1 GPU | 4000 | 26.28% | 75.97% | 26.28% | 38.86% |
82
+ | smart-test | Smart-Turn 3.2 CPU | 4000 | 92.50% | 92.59% | 92.50% | 92.50% |
83
+ | smart-test | Smart-Turn 3.2 GPU | 4000 | 93.88% | 93.89% | 93.88% | 93.87% |
84
+ | smart-test | Qwen3-Omni-30B-A3B-Instruct | 4000 | 69.40% | 80.43% | 69.40% | 66.35% |
85
+ | smart-test | Turn-Detection-Jev | 4000 | 96.78% | 96.87% | 96.78% | 96.81% |
86
+
87
+ Turn-Detection-Jev 获得了按标签规则构造的助手状态,其他音频模型未获得该状态,且服务和环境不同。因此表格只作描述性参照,不能据此将差值全部归因于微调或声称同等输入条件下的优势。
88
+
89
+ ### 评估协议与耗时
90
+
91
+ 公开集使用单声道、16 kHz、末尾最多 8 秒的 PCM16 WAV 音频;短音频保留实际长度。逐条串行预测,complete 分配到 listening,interrupt/backchannel 分配到 speaking;incomplete 按样本 ID 哈希分配一半到每个状态,状态种子为 4。助手状态为构造的场景元数据,未向麦克风音频混入助手播放声音。
92
+
93
+ | 完整公开集 | 平均逐样本耗时 | P50 | P90 | 串行 samples/s |
94
+ | --- | ---: | ---: | ---: | ---: |
95
+ | Easy-Turn(800) | 111.22 ms | 91.69 ms | 157.55 ms | 8.99 |
96
+ | Smart-Test(4,000) | 97.50 ms | 96.29 ms | 110.90 ms | 10.26 |
97
+
98
+ 本次计时包含本地文件读取/解码、音频预处理和模型推理,不经 HTTP;未剔除首次调用开销。samples/s=样本数×1000/累计逐样本耗时,不是整轮墙钟吞吐。分位数在排序后的 (N−1)×q 位置线性插值。未测量独立 GPU kernel 耗时或真实在线打断延迟。
99
+
100
+ 主页对比图中的 Easy-Turn 耗时仅对应 600 条共同样本,因此与上表 800 条的均值、分位数不同。其他模型的历史计时排除文件读取/解码,硬件和计时边界均不一致,不能作为同硬件速度排名。
LICENSE ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS
177
+
178
+ APPENDIX: How to apply the Apache License to your work.
179
+
180
+ To apply the Apache License to your work, attach the following
181
+ boilerplate notice, with the fields enclosed by brackets "[]"
182
+ replaced with your own identifying information. (Don't include
183
+ the brackets!) The text should be enclosed in the appropriate
184
+ comment syntax for the file format. We also recommend that a
185
+ file or class name and description of purpose be included on the
186
+ same "printed page" as the copyright notice for easier
187
+ identification within third-party archives.
188
+
189
+ Copyright [yyyy] [name of copyright owner]
190
+
191
+ Licensed under the Apache License, Version 2.0 (the "License");
192
+ you may not use this file except in compliance with the License.
193
+ You may obtain a copy of the License at
194
+
195
+ http://www.apache.org/licenses/LICENSE-2.0
196
+
197
+ Unless required by applicable law or agreed to in writing, software
198
+ distributed under the License is distributed on an "AS IS" BASIS,
199
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200
+ See the License for the specific language governing permissions and
201
+ limitations under the License.
NOTICE ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Turn-Detection-Jev
2
+
3
+ This release is a fine-tuned derivative of Qwen3-Omni-30B-A3B-Instruct,
4
+ developed and released by the Qwen team. Rights in the original model
5
+ remain with their respective holders.
6
+
7
+ Original model and upstream license declaration:
8
+ https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct
9
+
10
+ Official Qwen3-Omni source repository and Apache-2.0 license:
11
+ https://github.com/QwenLM/Qwen3-Omni
12
+ https://github.com/QwenLM/Qwen3-Omni/blob/main/LICENSE
13
+
14
+ Modification notice:
15
+ The base model was fine-tuned with LoRA for state-conditioned audio
16
+ turn completion, incomplete speech, interruption, backchannel and
17
+ background classification. Inference scores candidate answer tokens
18
+ through the language-model head. Historical experiment identifier:
19
+ Audio-Jev v3, checkpoint-1100, evaluated on 2026-11-02.
20
+
21
+ The fine-tuned weights in this release are distributed under the
22
+ Apache License, Version 2.0. See the accompanying LICENSE file.
23
+ Retain applicable original copyright and attribution notices when
24
+ redistributing base-model files or derived model files.
25
+
26
+ The Audio-Jev implementation uses ms-swift and includes work adapted
27
+ from OmniJev/OneJev (Apache-2.0):
28
+ https://github.com/modelscope/ms-swift
29
+ https://github.com/OmniJev/OneJev
30
+ If implementation code is distributed with the weights, retain its
31
+ applicable license and original notices as well.
32
+
33
+ Qwen and other project names are used only to identify the upstream
34
+ sources. This independent derivative release is not endorsed by the
35
+ upstream authors. Training datasets and third-party software retain
36
+ their respective licenses; this notice does not grant rights to
37
+ redistribute the original training audio.
38
+
39
+ Standalone packaging changes:
40
+ The frozen v3 prompt/schema/audio processing helpers were extracted from
41
+ ms-swift commit 96e8f8bdfe19b8ae3a2480797f09f69e4b33222b. Training-only
42
+ turn-policy code was omitted. The native Transformers/PEFT loader, CLI,
43
+ HTTP API, portable benchmark and release tooling were added for this
44
+ independent project. No ms-swift runtime dependency is required.
45
+ Dataset readers and the public-audio preprocessing profile were adapted
46
+ from the local audio-eval project. See SOURCES.json for file hashes.
README.md ADDED
@@ -0,0 +1,296 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen3-Omni-30B-A3B-Instruct
4
+ base_model_relation: finetune
5
+ library_name: transformers
6
+ pipeline_tag: audio-classification
7
+ tags:
8
+ - audio
9
+ - turn-taking
10
+ - end-of-turn-detection
11
+ - interruption-detection
12
+ - lora
13
+ - ms-swift
14
+ - custom-inference
15
+ datasets:
16
+ - ASLP-lab/Easy-Turn-Trainset
17
+ - pipecat-ai/smart-turn-data-v3.2-train
18
+ ---
19
+
20
+ # Turn-Detection-Jev
21
+
22
+ ### Audio turn completion, interruptions, and backchannels
23
+
24
+ **Fine-tuned Qwen3-Omni-30B-A3B-Instruct · Five semantic actions / seven state labels · Apache-2.0**
25
+
26
+ [GitHub code](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/tree/main) · [Model weights](https://huggingface.co/YOUR_HF_ORG/Turn-Detection-Jev) · [Quick start](#quick-start) · [License](LICENSE)
27
+
28
+ **Language:** English | [中文](README_zh.md)
29
+
30
+ Turn-Detection-Jev predicts when a speech assistant should consider responding, waiting, or handling a user interruption. It takes audio and the application's current assistant playback state, then scores the labels available in that state. It is fine-tuned from [Qwen3-Omni-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct).
31
+
32
+ ## Contents
33
+
34
+ - [Core features](#core-features)
35
+ - [Labels](#labels)
36
+ - [Benchmark results](#benchmark-results)
37
+ - [Metric comparison](#metric-comparison)
38
+ - [Client request latency](#client-request-latency)
39
+ - [Quick start](#quick-start)
40
+ - [Training and evaluation](#training-and-evaluation)
41
+ - [Limitations](#limitations)
42
+ - [License and acknowledgements](#license-and-acknowledgements)
43
+ - [Source code and feedback](#source-code-and-feedback)
44
+
45
+ ## Core features
46
+
47
+ | Capability | Description |
48
+ | --- | --- |
49
+ | Turn completion | Distinguishes finished speech, incomplete speech, and background when the assistant is silent |
50
+ | Interruption and backchannel | Distinguishes completed interruptions, brief feedback, unfinished speech, and background during playback |
51
+ | State-conditioned scoring | Uses three listening labels or four speaking labels according to caller-supplied playback state |
52
+ | Single-forward prediction | Uses Qwen3-Omni-30B-A3B-Instruct audio understanding and reads candidate answer-token logits in one forward pass, without generating explanations |
53
+ | Application integration | The GitHub runtime provides CLI inference and an HTTP API returning labels and probabilities |
54
+ | Open weights | Apache-2.0, retaining attribution to the Qwen3-Omni-30B-A3B-Instruct base model |
55
+
56
+ ## Labels
57
+
58
+ | Assistant playback | Output label | Meaning | Illustrative example (not an observed prediction) |
59
+ | --- | --- | --- | --- |
60
+ | Off | `listen_complete` | The user has finished and yielded the floor | “Check tomorrow's weather in Beijing.” |
61
+ | Off | `listen_incomplete` | The user is still speaking or holding the floor | “I'd like to ask about, um…” |
62
+ | Off | `listen_background` | Background unrelated to the target interaction | Ambient noise, bystanders, or playback echo |
63
+ | On | `speak_interrupt` | A completed request to stop, correct, or take the floor | “Stop a moment. I meant tomorrow.” |
64
+ | On | `speak_backchannel` | Finished brief feedback without requesting the floor | “Mm-hmm, right.” |
65
+ | On | `speak_incomplete` | The user has started but not finished an utterance | “Wait, what I wanted was…” |
66
+ | On | `speak_background` | Background unrelated to the target interaction | Ambient noise, bystanders, or playback echo |
67
+
68
+ The application must provide the real `assistant_speaking` state. The same words can express different intentions in context; these examples are illustrations, not measured model predictions.
69
+
70
+ ## Benchmark results
71
+
72
+ ### Public evaluation sets
73
+
74
+ Turn-Detection-Jev scores come from the complete native Transformers run `merged-20261111_051726-1672084` on 2026-11-11, using temperature T=1 with calibration disabled. Other models retain their historical results.
75
+
76
+ | Dataset | Samples | Correct | Accuracy | Semantic Macro-F1 |
77
+ | --- | ---: | ---: | ---: | ---: |
78
+ | Easy-Turn | 800 | 789 | **98.625%** | **99.00%** |
79
+ | Smart-Test-4000 | 4,000 | 3,871 | **96.775%** | **96.81%** |
80
+
81
+ **Evaluation conditions:** These are **state-conditioned single-window** results. Playback state was constructed using label rules and sample-ID hashes. Input audio was mono 16 kHz PCM16, retaining at most the final 8 seconds. Macro averages cover ground-truth-supported labels only. Easy-Turn covers five state labels and Smart-Test three; neither has background ground truth.
82
+
83
+ ### Comparison on shared samples
84
+
85
+ The main comparison uses original semantic labels: listen_complete→complete; listen_incomplete/speak_incomplete→incomplete; speak_interrupt→wait; speak_backchannel→backchannel; both background labels→invalid. Only labels are merged; predictions are retained, and options are not renormalized based on ground truth.
86
+
87
+ ## Metric comparison
88
+
89
+ | Dataset | Model | N | Accuracy | Macro precision | Macro recall | Macro-F1 |
90
+ | --- | --- | --- | --- | --- | --- | --- |
91
+ | easyturn | TurnSense 1.1 GPU | 600 | 96.17% | 96.19% | 96.17% | 96.17% |
92
+ | easyturn | Smart-Turn 3.2 CPU | 600 | 76.00% | 78.21% | 76.00% | 75.52% |
93
+ | easyturn | Smart-Turn 3.2 GPU | 600 | 77.17% | 78.26% | 77.17% | 76.94% |
94
+ | easyturn | Qwen3-Omni-30B-A3B-Instruct | 600 | 92.33% | 93.40% | 92.33% | 92.53% |
95
+ | easyturn | Turn-Detection-Jev | 600 | 98.17% | 98.33% | 98.17% | 98.25% |
96
+ | smart-test | TurnSense 1.1 GPU | 4000 | 26.28% | 75.97% | 26.28% | 38.86% |
97
+ | smart-test | Smart-Turn 3.2 CPU | 4000 | 92.50% | 92.59% | 92.50% | 92.50% |
98
+ | smart-test | Smart-Turn 3.2 GPU | 4000 | 93.88% | 93.89% | 93.88% | 93.87% |
99
+ | smart-test | Qwen3-Omni-30B-A3B-Instruct | 4000 | 69.40% | 80.43% | 69.40% | 66.35% |
100
+ | smart-test | Turn-Detection-Jev | 4000 | 96.78% | 96.87% | 96.78% | 96.81% |
101
+
102
+ ![Metrics on shared evaluation samples](image/01_metric_comparison.png)
103
+
104
+ Accuracy is computed over all shared samples. Precision, recall, and F1 are equally averaged over the two ground-truth target classes, complete and incomplete. Predictions outside those targets remain errors and are not renormalized.
105
+
106
+ Turn-Detection-Jev received constructed assistant-state metadata that the other audio models did not receive. Services and execution environments also differ. These figures compare the current Jev run with historical baselines; they are not a controlled comparison with identical information. The TurnSense figures here come from our evaluation, not its official model card.
107
+
108
+ ### Client request latency
109
+
110
+ | Dataset | Model | N | Mean ms | P50 ms | P90 ms | Serial samples/s |
111
+ | --- | --- | --- | --- | --- | --- | --- |
112
+ | easyturn | TurnSense 1.1 GPU | 600 | 638.60 | 648.29 | 657.59 | 1.57 |
113
+ | easyturn | Smart-Turn 3.2 CPU | 600 | 64.91 | 47.66 | 54.51 | 15.41 |
114
+ | easyturn | Smart-Turn 3.2 GPU | 600 | 5.36 | 4.18 | 5.26 | 186.68 |
115
+ | easyturn | Qwen3-Omni-30B-A3B-Instruct | 600 | 222.09 | 226.87 | 247.68 | 4.50 |
116
+ | easyturn | Turn-Detection-Jev | 600 | 120.38 | 94.68 | 160.56 | 8.31 |
117
+ | smart-test | TurnSense 1.1 GPU | 4000 | 543.53 | 649.01 | 659.98 | 1.84 |
118
+ | smart-test | Smart-Turn 3.2 CPU | 4000 | 165.56 | 192.63 | 296.12 | 6.04 |
119
+ | smart-test | Smart-Turn 3.2 GPU | 4000 | 4.07 | 3.42 | 4.63 | 245.65 |
120
+ | smart-test | Qwen3-Omni-30B-A3B-Instruct | 4000 | 229.76 | 221.62 | 275.89 | 4.35 |
121
+ | smart-test | Turn-Detection-Jev | 4000 | 97.50 | 96.29 | 110.90 | 10.26 |
122
+
123
+ ![Client request latency and serial throughput](image/02_latency_comparison.png)
124
+
125
+ Turn-Detection-Jev latency comes from the native Transformers run above and includes local audio reading/decoding, preprocessing and inference. First-call overhead is retained; independent GPU kernel time was not measured. Other models retain historical timing.
126
+
127
+ The chart uses logarithmic axes. Historical ONNX runs used warmup=20 and remote runs warmup=0. ONNX timing includes mixing, resampling, features and inference; remote timing additionally includes encoding, HTTP, queueing and parsing. Both historical paths exclude file reading/decoding, whereas this Jev run includes it. Hardware, models and execution environments also differ, so this is not a controlled speed ranking.
128
+
129
+ Table samples/s equals N × 1000 / cumulative per-sample latency; it is neither whole-run wall-clock throughput nor concurrent capacity. Independent `gpu_latency_ms` and `inference_latency_ms` were not recorded in this run.
130
+
131
+ ## Quick start
132
+
133
+ This Hugging Face repository contains model artifacts and documentation. Inference, serving, evaluation tools, and test-set instructions are maintained in the linked GitHub repository.
134
+
135
+ Copy the complete project to the server, or obtain its source first (Git does not include the test WAV files):
136
+
137
+ ```bash
138
+ git clone https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev.git
139
+ cd Turn-Detection-Jev
140
+ git checkout main
141
+ ```
142
+
143
+ ### 1. Install dependencies
144
+
145
+ Activate an isolated Python >= 3.10 environment and install CUDA-enabled PyTorch suitable for the server first. From the project root, run:
146
+
147
+ ```bash
148
+ bash scripts/01_install.sh
149
+ ```
150
+
151
+ Alternatively, run the equivalent commands directly (choose either the script or these commands):
152
+
153
+ ```bash
154
+ python -m pip install -e '.[datasets]'
155
+ ```
156
+
157
+ This runs `python -m pip install -e '.[datasets]'` and prints the PyTorch version and CUDA availability. Dependencies are declared in `pyproject.toml`; the datasets extra installs PyArrow for raw Parquet conversion in step 2; merged weights do not require the PEFT/adapter extra. Confirm `CUDA available: True` before continuing. All three scripts use the current environment's `python`, falling back to `python3`, or an explicit `PYTHON=/absolute/path/to/python`. Use the same environment for all steps.
158
+
159
+ ### 2. Convert and validate the original test sets
160
+
161
+ Before step 2, keep this layout in the code repository when using default paths:
162
+
163
+ ```text
164
+ Turn-Detection-Jev/
165
+ ├── datasets/
166
+ │ ├── README.md # Dataset instructions; recommended
167
+ │ ├── Easy-Turn-Testset/ # Complete original Easy-Turn test set
168
+ │ ├── smart-turn-data-v3.2-test/ # Complete original Smart-Turn v3.2 test set
169
+ │ ├── easyturn/ # Documentation/licenses copied during export
170
+ │ │ ├── LICENSE
171
+ │ │ ├── README.md
172
+ │ │ └── SOURCE_README.md
173
+ │ ├── smart-test-4000/ # Documentation/licenses copied during export
174
+ │ │ ├── LICENSE
175
+ │ │ ├── README.md
176
+ │ │ └── SOURCE_README.md
177
+ │ └── v3-public/ # Generated by step 2; do not create an empty directory
178
+ │ ├── easyturn/ # audio/, test.jsonl, summary.json, docs/licenses
179
+ │ └── smart-test-4000/ # audio/, test.jsonl, summary.json, docs/licenses
180
+ └── results/
181
+ └── reference/ # Frozen IDs, labels, states and audio hashes
182
+ ├── easyturn.predictions.jsonl
183
+ └── smart-test-4000.predictions.jsonl
184
+ ```
185
+
186
+ Keep the original upstream layouts and audio. See the code repository's `datasets/README.md` for required lists and Parquet shards. Old `audio/`, `test.jsonl` and `summary.json` files directly under `datasets/easyturn/` and `datasets/smart-test-4000/` are not inputs to the current steps 2 and 3 and do not need uploading; retain the three documentation/license files shown in each directory.
187
+
188
+ Step 3 reads the generated `datasets/v3-public/`; keep its audio, manifests and accompanying files together. Step 2 validates and reuses a complete existing export without requiring the original datasets. Rebuilding requires the originals, documentation/licenses and `results/reference/` again. Keep the remaining source code and configuration files as usual.
189
+
190
+ With the default layout above, run:
191
+
192
+ ```bash
193
+ bash scripts/02_validate_data.sh
194
+ ```
195
+
196
+ Use the original datasets on the evaluation server (paths may contain spaces):
197
+
198
+ ```bash
199
+ EASYTURN_DIR=/path/to/Easy-Turn-Testset \
200
+ SMART_TEST_DIR=/path/to/smart-turn-data-v3.2-test \
201
+ bash scripts/02_validate_data.sh
202
+ ```
203
+
204
+ Without these variables, sources default to `datasets/Easy-Turn-Testset` and `datasets/smart-turn-data-v3.2-test` inside this project. The default input is the full original Smart-Turn v3.2 test snapshot; conversion selects the frozen historical 4000 IDs. You can also point `SMART_TEST_DIR` to an existing 4000-row raw Parquet subset.
205
+
206
+ Step 2 converts to `datasets/v3-public/` and validates paths, SHA256 hashes, audio format and label/state consistency for **800** Easy-Turn and **4000** Smart-Test samples. Audio is mono 16 kHz PCM16 WAV, keeping at most the last 8 seconds. No model is loaded and sources are unchanged. Later runs fully validate and reuse existing output. Damaged or incomplete output causes an error; use a new `DATASETS_DIR` to rebuild.
207
+
208
+ Git excludes `datasets/**/audio/`. Convert from the originals on the server without uploading a local audio export. Step 2 does not download the originals automatically.
209
+
210
+ To change the output directory, set the same `DATASETS_DIR` for steps 2 and 3:
211
+
212
+ ```bash
213
+ EASYTURN_DIR=/path/to/Easy-Turn-Testset \
214
+ SMART_TEST_DIR=/path/to/smart-turn-data-v3.2-test \
215
+ DATASETS_DIR=datasets/v3-public-new bash scripts/02_validate_data.sh
216
+ DATASETS_DIR=datasets/v3-public-new bash scripts/03_evaluate.sh
217
+ ```
218
+
219
+ Equivalent direct commands (choose either these or the script):
220
+
221
+ ```bash
222
+ python scripts/prepare_v3_testsets.py \
223
+ --easyturn /path/to/Easy-Turn-Testset \
224
+ --smart-test /path/to/smart-turn-data-v3.2-test \
225
+ --output datasets/v3-public --reuse-existing
226
+ python evaluate.py --manifest datasets/v3-public/easyturn/test.jsonl --validate-only
227
+ python evaluate.py --manifest datasets/v3-public/smart-test-4000/test.jsonl --validate-only
228
+ ```
229
+
230
+ ### 3. Evaluate merged weights
231
+
232
+ Place the full model in `models/Turn-Detection-Jev/`. Include `config.json`, all weight shards and their index, processor/tokenizer files and the original `scoring_head.json`. Do not leave an adapter's `adapter_config.json` in this directory: it selects the adapter loading path.
233
+
234
+ ```bash
235
+ bash scripts/03_evaluate.sh
236
+
237
+ # Optional: specify a model directory and a new output directory.
238
+ # Relative paths are resolved from the project root.
239
+ bash scripts/03_evaluate.sh models/Turn-Detection-Jev results/runs/merged-check
240
+ ```
241
+
242
+ Alternatively, run the equivalent commands directly (choose either the script or these commands):
243
+
244
+ ```bash
245
+ RUN_OUTPUT="results/runs/merged-$(date +%Y%m%d_%H%M%S)-$$"
246
+ python evaluate.py --model models/Turn-Detection-Jev \
247
+ --device cuda:0 --dtype bfloat16 --attn-implementation sdpa \
248
+ --calibration hf/calibration.json \
249
+ --manifest datasets/v3-public/easyturn/test.jsonl --out "$RUN_OUTPUT/easyturn" \
250
+ --manifest datasets/v3-public/smart-test-4000/test.jsonl --out "$RUN_OUTPUT/smart-test-4000"
251
+ ```
252
+
253
+ By default this reads the converted data from `datasets/v3-public/` validated in step 2. This starts one `evaluate.py` process, loads the model once and sequentially evaluates all 800 Easy-Turn and 4000 Smart-Test samples, without `--limit`. It uses native Transformers/PyTorch, one GPU (`cuda:0`), BF16, SDPA and temperature calibration from `hf/calibration.json`. No HTTP server, vLLM, separate base-model download or LoRA weights are needed. Select another GPU with `DEVICE=cuda:1 bash scripts/03_evaluate.sh`; the selected GPU must have enough memory for inference.
254
+
255
+ The script now loads `hf/calibration.json` by default: listening uses `choice:3` (T≈1.01754) and speaking uses `choice:4` (T≈1.37385). The report records `provenance.calibration_enabled=true`. To select another file:
256
+
257
+ ```bash
258
+ CALIBRATION=/path/to/calibration.json bash scripts/03_evaluate.sh /path/to/model
259
+ ```
260
+
261
+ The published 2026-11-11 scores above came from a run with calibration disabled and have not been replaced with a calibrated rerun. Use this command to match that T=1 setting:
262
+
263
+ ```bash
264
+ CALIBRATION=none bash scripts/03_evaluate.sh /path/to/model
265
+ ```
266
+
267
+ Temperature scaling adjusts probabilities and confidence. Positive temperature preserves label ordering mathematically and does not guarantee an improvement in Accuracy/F1.
268
+
269
+ Each run defaults to a new `results/runs/merged-TIMESTAMP-PID/` directory with per-dataset `predictions.jsonl` and `report.json` files. Inspect `classification.accuracy`, `classification.macro_f1`, `semantic.macro_f1` and `is_full_run=true`. The main benchmark table reports semantic Macro-F1. A specified output directory must not already exist; scripts stop on errors.
270
+
271
+ These are single-window results conditioned on constructed `assistant_speaking` state. Published Jev scores now use the complete native Transformers run above; historical ms-swift results remain in the code repository under `results/reference/`.
272
+
273
+
274
+ ## Training and evaluation
275
+
276
+ The model was fine-tuned with LoRA rank 32, alpha 64, learning rate 5e-5, and a three-epoch budget. The processed pool contains 156,925 samples from Easy-Turn, Smart-Turn and MUSAN noise, with 141,232 assigned to training. On the separate internal test set of 3,923 samples, accuracy was 97.04% and seven-label Macro-F1 was 97.87%.
277
+
278
+ See [EVALUATION.md](EVALUATION.md) for data distributions, label recalls, experiment provenance and timing boundaries.
279
+
280
+ ## Limitations
281
+
282
+ - Public sets do not assess background recall. Smart-Test's speaking group contains incomplete speech only, so its 99.90% accuracy does not establish interruption/backchannel discrimination.
283
+ - Performance is conditional on correct external playback state. These results do not establish streaming event latency or robustness to deployment devices, echo, noise and accents.
284
+ - This release requires a Qwen3-Omni-30B-A3B-Instruct GPU environment. CPU, ONNX and quantized deployments have not been benchmarked.
285
+
286
+ ## License and acknowledgements
287
+
288
+ Fine-tuned weights are released under [Apache License 2.0](LICENSE), permitting no-charge research, personal and commercial use, modification and redistribution subject to its terms.
289
+
290
+ The base model was developed and released by the Qwen team: [Qwen3-Omni-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct). Its official model page declares Apache-2.0; the official license text is available in [Qwen3-Omni/LICENSE](https://github.com/QwenLM/Qwen3-Omni/blob/main/LICENSE). Base-model attribution and fine-tuning modifications are recorded in [NOTICE](NOTICE).
291
+
292
+ Thanks to Qwen, ms-swift, OmniJev/OneJev, Easy-Turn, Smart-Turn and MUSAN. Datasets and dependencies retain their own licenses. This independent derivative is not endorsed by the upstream authors.
293
+
294
+ ## Source code and feedback
295
+
296
+ [GitHub source](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/tree/main) · [GitHub Issues](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/issues)
README_zh.md ADDED
@@ -0,0 +1,296 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen3-Omni-30B-A3B-Instruct
4
+ base_model_relation: finetune
5
+ library_name: transformers
6
+ pipeline_tag: audio-classification
7
+ tags:
8
+ - audio
9
+ - turn-taking
10
+ - end-of-turn-detection
11
+ - interruption-detection
12
+ - lora
13
+ - ms-swift
14
+ - custom-inference
15
+ datasets:
16
+ - ASLP-lab/Easy-Turn-Trainset
17
+ - pipecat-ai/smart-turn-data-v3.2-train
18
+ ---
19
+
20
+ # Turn-Detection-Jev
21
+
22
+ ### 面向语音对话的轮次结束、打断与附和判断
23
+
24
+ **Qwen3-Omni-30B-A3B-Instruct 微调 · 五种语义 / 七个状态标签 · Apache-2.0**
25
+
26
+ [GitHub 源码](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/tree/main) · [模型权重](https://huggingface.co/YOUR_HF_ORG/Turn-Detection-Jev) · [快速开始](#快速开始) · [许可证](LICENSE)
27
+
28
+ **语言:** [English](README.md) | 中文
29
+
30
+
31
+ **Turn-Detection-Jev** 用于判断语音对话中何时接话、何时等待,以及助手播放过程中用户是否希望打断。模型同时接收音频与应用提供的助手播放状态,输出对应场景的类别及概率,可用于语音助手、客服和语音对话系统的轮次控制研究与集成。
32
+
33
+ Turn-Detection-Jev is a state-conditioned audio model for turn completion, interruptions, backchannels, incomplete speech, and background audio, fine-tuned from Qwen3-Omni-30B-A3B-Instruct.
34
+
35
+ ## 目录
36
+
37
+ - [核心特性](#核心特性)
38
+ - [模型如何判断](#模型如何判断)
39
+ - [基准测试结果](#基准测试结果)
40
+ - [快速开始](#快速开始)
41
+ - [训练与评测说明](#训练与评测说明)
42
+ - [使用边界](#使用边界)
43
+ - [许可证与致谢](#许可证与致谢)
44
+
45
+ ## 核心特性
46
+
47
+ | 能力 | 说明 |
48
+ | --- | --- |
49
+ | 轮次结束判断 | 在助手未播放时,区分用户说完、尚未说完和背景声音 |
50
+ | 打断与附和判断 | 在助手播放时,区分已完成的打断请求、简短附和、未完成表达和背景声音 |
51
+ | 状态条件决策 | 依据应用提供的 `assistant_speaking`,在三种或四种候选标签中评分 |
52
+ | 基于音频评分 | 使用 Qwen3-Omni-30B-A3B-Instruct 的音频理解能力,通过一次前向计算读取选项 logits,无需生成解释文本 |
53
+ | 可接入应用 | 通过 HTTP 返回预测类别与候选概率,供应用的轮次控制逻辑使用 |
54
+ | 免费开放使用 | 微调权重采用 Apache-2.0,保留 Qwen3-Omni-30B-A3B-Instruct 基座归属与许可声明 |
55
+
56
+ ## 模型如何判断
57
+
58
+ 模型包含 complete、incomplete、interrupt、backchannel、background 五种语义。结合助手是否正在播放,形成七个监督标签。
59
+
60
+ | 助手状态 | 输出标签 | 含义 | 场景示例(仅用于说明) |
61
+ | --- | --- | --- | --- |
62
+ | 未播放 | `listen_complete` | 用户完成表达,交出话轮 | “帮我查一下明天北京的天气。” |
63
+ | 未播放 | `listen_incomplete` | 用户尚未完成,仍持有话轮 | “我想问一下,那个……” |
64
+ | 未播放 | `listen_background` | 与当前交互无关的背景声音 | 环境噪声、旁人声音或播放回声 |
65
+ | 播放中 | `speak_interrupt` | 用户完成了停止、纠正或接管话轮的表达 | “先停一下,我说的是明天。” |
66
+ | 播放中 | `speak_backchannel` | 用户完成简短反馈,无意接管话轮 | “嗯,对。” |
67
+ | 播放中 | `speak_incomplete` | 用户已经开始表达,但尚未完成 | “等一下,我是想……” |
68
+ | 播放中 | `speak_background` | 与当前交互无关的背景声音 | 环境噪声、旁人声音或播放回声 |
69
+
70
+ `assistant_speaking` 应由应用的真实播放状态提供。同一句话在不同语境中可能有不同意图;上述例子不代表实测预测结果。
71
+
72
+ ## 基准测试结果
73
+
74
+ ### 公开测试集
75
+
76
+ Turn-Detection-Jev 成绩来自 2026-11-11 的完整原生 Transformers 评估。
77
+
78
+ | 数据集 | 样本数 | 正确数 | Accuracy | 语义 Macro-F1 |
79
+ | --- | ---: | ---: | ---: | ---: |
80
+ | Easy-Turn | 800 | 789 | **98.625%** | **99.00%** |
81
+ | Smart-Test-4000 | 4,000 | 3,871 | **96.775%** | **96.81%** |
82
+
83
+ **评估条件**:以上为已知助手状态的单窗口分类结果。状态由标签规则与样本 ID 构造,输入为单声道 16 kHz、末尾最多 8 秒音频;不代表纯音频分类或在线系统表现。Easy-Turn 覆盖五个状态标签,Smart-Test 覆盖三个,均没有 background 真值。语义 Macro-F1 只计有真值支持的类别。
84
+
85
+ ### 共同样本上的模型参照
86
+
87
+ 主比较统一使用原语义标签:listen_complete→complete,listen_incomplete/speak_incomplete→incomplete,speak_interrupt→wait,speak_backchannel→backchannel,两种 background→invalid。只合并标签,不删除预测或按真值重新选项归一化。
88
+
89
+ ## 指标对比
90
+
91
+ | 数据集 | 模型 | N | 准确率 | 精准率(宏平均) | 召回率(宏平均) | F1(宏平均) |
92
+ | --- | --- | --- | --- | --- | --- | --- |
93
+ | easyturn | TurnSense 1.1 GPU | 600 | 96.17% | 96.19% | 96.17% | 96.17% |
94
+ | easyturn | Smart-Turn 3.2 CPU | 600 | 76.00% | 78.21% | 76.00% | 75.52% |
95
+ | easyturn | Smart-Turn 3.2 GPU | 600 | 77.17% | 78.26% | 77.17% | 76.94% |
96
+ | easyturn | Qwen3-Omni-30B-A3B-Instruct | 600 | 92.33% | 93.40% | 92.33% | 92.53% |
97
+ | easyturn | Turn-Detection-Jev | 600 | 98.17% | 98.33% | 98.17% | 98.25% |
98
+ | smart-test | TurnSense 1.1 GPU | 4000 | 26.28% | 75.97% | 26.28% | 38.86% |
99
+ | smart-test | Smart-Turn 3.2 CPU | 4000 | 92.50% | 92.59% | 92.50% | 92.50% |
100
+ | smart-test | Smart-Turn 3.2 GPU | 4000 | 93.88% | 93.89% | 93.88% | 93.87% |
101
+ | smart-test | Qwen3-Omni-30B-A3B-Instruct | 4000 | 69.40% | 80.43% | 69.40% | 66.35% |
102
+ | smart-test | Turn-Detection-Jev | 4000 | 96.78% | 96.87% | 96.78% | 96.81% |
103
+
104
+ ![共同样本指标对比](image/01_metric_comparison.png)
105
+
106
+ 准确率按全部共同样本计算;精准率、召回率、F1 均对 complete 和 incomplete 两个真实目标类等权平均。所有目标外预测仍计入错误,不重新归一化。
107
+
108
+ Turn-Detection-Jev 额外获得了构造的助手状态,其他音频模型未获得该输入;各服务和运行环境也不同。因此这些数值为本次 Jev 与其他模型历史实现的描述性参照,不能据此声称同等输入下的领先或将差值全部归因于微调。这里的 TurnSense 数据来自本项目的评估,不是 TurnSense 官方页面公布的成绩。
109
+
110
+ ### 调用耗时
111
+
112
+ | 数据集 | 模型 | N | 均值 ms | P50 ms | P90 ms | 串行 samples/s |
113
+ | --- | --- | --- | --- | --- | --- | --- |
114
+ | easyturn | TurnSense 1.1 GPU | 600 | 638.60 | 648.29 | 657.59 | 1.57 |
115
+ | easyturn | Smart-Turn 3.2 CPU | 600 | 64.91 | 47.66 | 54.51 | 15.41 |
116
+ | easyturn | Smart-Turn 3.2 GPU | 600 | 5.36 | 4.18 | 5.26 | 186.68 |
117
+ | easyturn | Qwen3-Omni-30B-A3B-Instruct | 600 | 222.09 | 226.87 | 247.68 | 4.50 |
118
+ | easyturn | Turn-Detection-Jev | 600 | 120.38 | 94.68 | 160.56 | 8.31 |
119
+ | smart-test | TurnSense 1.1 GPU | 4000 | 543.53 | 649.01 | 659.98 | 1.84 |
120
+ | smart-test | Smart-Turn 3.2 CPU | 4000 | 165.56 | 192.63 | 296.12 | 6.04 |
121
+ | smart-test | Smart-Turn 3.2 GPU | 4000 | 4.07 | 3.42 | 4.63 | 245.65 |
122
+ | smart-test | Qwen3-Omni-30B-A3B-Instruct | 4000 | 229.76 | 221.62 | 275.89 | 4.35 |
123
+ | smart-test | Turn-Detection-Jev | 4000 | 97.50 | 96.29 | 110.90 | 10.26 |
124
+
125
+ ![调用耗时与串行吞吐对比](image/02_latency_comparison.png)
126
+
127
+ Turn-Detection-Jev 耗时来自上述原生 Transformers 运行,含本地音频读取/解码、预处理和推理;未剔除首次调用开销,未测量独立 GPU kernel 耗时。其他模型仍为历史计时。
128
+
129
+ 图表使用对数轴。其他模型的 ONNX 运行 warmup=20,历史远端运行 warmup=0;历史 ONNX 计时含混音、重采样、特征和推理,远端还含编码、HTTP、排队及响应解析,均不含文件读取/解码。Turn-Detection-Jev 本次计时包含读取/解码,且硬件、模型和运行环境不同,本图不能用于同硬件速度排名。
130
+
131
+ 表中 samples/s=样本数×1000/累计逐样本耗时,不是整轮墙钟吞吐,也不表示并发容量。未在本次输出中记录独立 `gpu_latency_ms` 或 `inference_latency_ms`。
132
+
133
+ ## 快速开始
134
+
135
+ 本 Hugging Face 仓库提供模型权重、模型卡和配套评分/校准文件。推理、HTTP 服务、评估工具和测试集说明在 [GitHub 源码仓库](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/tree/main)。
136
+
137
+ 在远程服务器复制完整项目,或先获取源码(代码仓库不包含测试 WAV):
138
+
139
+ ```bash
140
+ git clone https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev.git
141
+ cd Turn-Detection-Jev
142
+ git checkout main
143
+ ```
144
+
145
+ ### 1. 安装依赖
146
+
147
+ 先激活 Python >= 3.10 的独立环境,并按服务器 GPU/CUDA 环境安装相应的 CUDA 版 PyTorch。在项目根目录执行:
148
+
149
+ ```bash
150
+ bash scripts/01_install.sh
151
+ ```
152
+
153
+ 也可直接执行以下等价命令(与脚本二选一):
154
+
155
+ ```bash
156
+ python -m pip install -e '.[datasets]'
157
+ ```
158
+
159
+ 脚本执行 `python -m pip install -e '.[datasets]'`,随后显示 PyTorch 版本和 CUDA 是否可用。依赖由 `pyproject.toml` 管理;`datasets` 扩展安装原始 Parquet 转换所需的 PyArrow;完整合并模型不需要 PEFT/adapter 扩展。确认 `CUDA available: True` 后继续。三个脚本均使用当前环境的 `python`(不存在时用 `python3`),也可通过 `PYTHON=/绝对路径/python` 指定解释器;后续步骤应使用同一环境。
160
+
161
+ ### 2. 转换原始数据并校验
162
+
163
+ 按默认路径执行第二步前,项目中应保留以下文件和目录:
164
+
165
+ ```text
166
+ Turn-Detection-Jev/
167
+ ├── datasets/
168
+ │ ├── README.md # 数据说明,建议保留
169
+ │ ├── Easy-Turn-Testset/ # 完整原始 Easy-Turn 测试集
170
+ │ ├── smart-turn-data-v3.2-test/ # 完整原始 Smart-Turn v3.2 测试集
171
+ │ ├── easyturn/ # 转换时复制的说明和许可文件
172
+ │ │ ├── LICENSE
173
+ │ │ ├── README.md
174
+ │ │ └── SOURCE_README.md
175
+ │ ├── smart-test-4000/ # 转换时复制的说明和许可文件
176
+ │ │ ├── LICENSE
177
+ │ │ ├── README.md
178
+ │ │ └── SOURCE_README.md
179
+ │ └── v3-public/ # 第二步成功后自动生成,首次不要创建空目录
180
+ │ ├── easyturn/ # audio/、test.jsonl、summary.json 和说明/许可
181
+ │ └── smart-test-4000/ # audio/、test.jsonl、summary.json 和说明/许可
182
+ └── results/
183
+ └── reference/ # 固定样本 ID、标签、状态和音频哈希所需
184
+ ├── easyturn.predictions.jsonl
185
+ └── smart-test-4000.predictions.jsonl
186
+ ```
187
+
188
+ 两个原始数据集须保留上游目录结构及音频内容;具体清单和 Parquet 分片要求见代码仓库的 `datasets/README.md`。顶层 `datasets/easyturn/`、`datasets/smart-test-4000/` 中的旧 `audio/`、`test.jsonl`、`summary.json` 不是当前第二、三步的输入,无需上传;上述各自的三个说明/许可文件仍需保留。
189
+
190
+ 第三步默认只读取新生成的 `datasets/v3-public/`,需完整保留其中的音频、清单和配套文件。已有完整转换结果时,第二步校验并复用该目录,不再要求原始数据目录存在;重新转换时仍需原始数据、说明/许可文件和 `results/reference/`。其余项目源码和配置文件照常保留。
191
+
192
+ 按上述默认目录放置后,直接执行:
193
+
194
+ ```bash
195
+ bash scripts/02_validate_data.sh
196
+ ```
197
+
198
+ 服务器上已有原始测试集时,指定实际目录执行(路径可包含空格):
199
+
200
+ ```bash
201
+ EASYTURN_DIR=/path/to/Easy-Turn-Testset \
202
+ SMART_TEST_DIR=/path/to/smart-turn-data-v3.2-test \
203
+ bash scripts/02_validate_data.sh
204
+ ```
205
+
206
+ 不设置环境变量时,默认从本项目的 `datasets/Easy-Turn-Testset` 和 `datasets/smart-turn-data-v3.2-test` 读取原始数据。默认读取完整的原始 Smart-Turn v3.2 测试快照,转换时仅选取冻结的历史 4000 个 ID;也可通过 `SMART_TEST_DIR` 指定已有的 4000 条原始 Parquet 子集。
207
+
208
+ 第二步自动转换到 `datasets/v3-public/`,校验 Easy-Turn **800** 条和 Smart-Test **4000** 条的路径、SHA256、音频格式及标签/状态。音频转换为单声道 16 kHz、末尾最多 8 秒 PCM16 WAV,不加载模型、不修改原始数据。后续运行会完整校验并复用已有输出,无需再次转换;已有输出损坏或不完整时会报错,请用新的 `DATASETS_DIR` 重新转换。
209
+
210
+ Git 不包含 `datasets/**/audio/`;无需从本地上传导出的测试音频,使用服务器上的原始数据转换即可。第二步不会自动下载原始数据。
211
+
212
+ 自定义输出目录时,第二步和第三步使用相同的 `DATASETS_DIR`:
213
+
214
+ ```bash
215
+ EASYTURN_DIR=/path/to/Easy-Turn-Testset \
216
+ SMART_TEST_DIR=/path/to/smart-turn-data-v3.2-test \
217
+ DATASETS_DIR=datasets/v3-public-new bash scripts/02_validate_data.sh
218
+ DATASETS_DIR=datasets/v3-public-new bash scripts/03_evaluate.sh
219
+ ```
220
+
221
+ 也可直接执行以下等价命令(与脚本二选一):
222
+
223
+ ```bash
224
+ python scripts/prepare_v3_testsets.py \
225
+ --easyturn /path/to/Easy-Turn-Testset \
226
+ --smart-test /path/to/smart-turn-data-v3.2-test \
227
+ --output datasets/v3-public --reuse-existing
228
+ python evaluate.py --manifest datasets/v3-public/easyturn/test.jsonl --validate-only
229
+ python evaluate.py --manifest datasets/v3-public/smart-test-4000/test.jsonl --validate-only
230
+ ```
231
+
232
+ ### 3. 评估完整合并模型
233
+
234
+ 经 LoRA 合并后的完整模型(或从本 HF 仓库下载的完整模型)放入 `models/Turn-Detection-Jev/`,包含 `config.json`、全部权重分片及索引、processor/tokenizer 配套文件,并保留原始 `scoring_head.json`。不要混入 adapter 的 `adapter_config.json`,否则加载器会进入 adapter 分支。
235
+
236
+ ```bash
237
+ bash scripts/03_evaluate.sh
238
+
239
+ # 可选:指定模型目录及一个尚不存在的结果目录(相对路径以项目根目录为准)。
240
+ bash scripts/03_evaluate.sh models/Turn-Detection-Jev results/runs/merged-check
241
+ ```
242
+
243
+ 也可直接执行以下等价命令(与脚本二选一):
244
+
245
+ ```bash
246
+ RUN_OUTPUT="results/runs/merged-$(date +%Y%m%d_%H%M%S)-$$"
247
+ python evaluate.py --model models/Turn-Detection-Jev \
248
+ --device cuda:0 --dtype bfloat16 --attn-implementation sdpa \
249
+ --calibration hf/calibration.json \
250
+ --manifest datasets/v3-public/easyturn/test.jsonl --out "$RUN_OUTPUT/easyturn" \
251
+ --manifest datasets/v3-public/smart-test-4000/test.jsonl --out "$RUN_OUTPUT/smart-test-4000"
252
+ ```
253
+
254
+ 脚本默认读取第二步生成并校验的 `datasets/v3-public/`,在同一个 `evaluate.py` 进程中只加载一次模型,依次完整评估 Easy-Turn 800 条与 Smart-Test 4000 条,不设置 `--limit`。使用原生 Transformers/PyTorch、单卡 `cuda:0`、BF16、SDPA,默认加载 `hf/calibration.json` 进行温度校准,无需启动 HTTP 服务或 vLLM,也无需另行下载基座/LoRA 权重。可用 `DEVICE=cuda:1 bash scripts/03_evaluate.sh` 指定 GPU;该卡需要容纳完整推理所需显存。
255
+
256
+ 脚本现默认加载 `hf/calibration.json`:listening 使用 `choice:3`(T≈1.01754),speaking 使用 `choice:4`(T≈1.37385)。报告记录 `provenance.calibration_enabled=true`。也可指定其他校准文件:
257
+
258
+ ```bash
259
+ CALIBRATION=/path/to/calibration.json bash scripts/03_evaluate.sh /path/to/model
260
+ ```
261
+
262
+ 上述 2026-11-11 公布成绩来自未启用 calibration 的完整运行,尚未替换为校准后的重跑结果。若要使用相同的 T=1 设置,显式关闭校准:
263
+
264
+ ```bash
265
+ CALIBRATION=none bash scripts/03_evaluate.sh /path/to/model
266
+ ```
267
+
268
+ 温度校准调整概率和置信度;正温度在数学上保持类别排序,不保证提高 Accuracy/F1。
269
+
270
+ 默认结果保存到自动生成的 `results/runs/merged-时间戳-进程号/`,两个数据集各有 `predictions.jsonl` 和 `report.json`。查看 `classification.accuracy`、`classification.macro_f1`、`semantic.macro_f1`,并确认 `is_full_run=true`;主页主表使用语义 Macro-F1。指定的结果目录必须不存在,脚本遇到错误会停止。
271
+
272
+ 以上是已构造 `assistant_speaking` 状态条件下的单窗口评估。主页 Jev 成绩已更新为上述原生 Transformers 完整运行;历史 ms-swift 成绩保留于代码仓库的 `results/reference/`。
273
+
274
+ ## 训练与评测说明
275
+
276
+ 模型基于 Qwen3-Omni-30B-A3B-Instruct,经 LoRA 微调:rank 32、alpha 64、学习率 5e-5、训练预算 3 epochs。数据池共 156,925 条,来自 Easy-Turn、Smart-Turn 和 MUSAN noise;实际训练划分为 141,232 条。内部 3923 条 test 的 Accuracy 为 97.04%,七类 Macro-F1 为 97.87%。
277
+
278
+ 详细标签召回率、来源、实验版本和计时边界见 [EVALUATION.md](EVALUATION.md)。
279
+
280
+ ## 使用边界
281
+
282
+ - 公开集无 background 真值;Smart-Test speaking 组只有 incomplete,不能验证打断和附和的完整区分能力。
283
+ - 性能依赖正确的外部播放状态;当前成绩不代表流式在线打断延迟或跨设备、噪声、回声下的可靠性。
284
+ - 需要可运行 Qwen3-Omni-30B-A3B-Instruct 的 GPU 环境;当前未提供 CPU、ONNX 或量化后端的验证结果。
285
+
286
+ ## 许可证与致谢
287
+
288
+ 微调权重采用 [Apache License 2.0](LICENSE),可在遵守条款的前提下免费用于研究、个人和商业应用,以及修改和再分发。
289
+
290
+ 基座由 Qwen 团队发布:[Qwen3-Omni-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct) 的官方模型页标识为 Apache-2.0,许可全文参见 [Qwen3-Omni 官方 LICENSE](https://github.com/QwenLM/Qwen3-Omni/blob/main/LICENSE)。本发布保留基座归属与微调修改说明,见 [NOTICE](NOTICE)。
291
+
292
+ 感谢 Qwen、ms-swift、OmniJev/OneJev、Easy-Turn、Smart-Turn 和 MUSAN。各数据集和依赖保留其自身许可;独立衍生模型不代表上游作者背书。
293
+
294
+ ## 源码与反馈
295
+
296
+ 推理代码、测试集准备与评估:[GitHub](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/tree/main)。问题反馈:[GitHub Issues](https://github.com/YOUR_GITHUB_ORG/Turn-Detection-Jev/issues)。
calibration.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "temperatures": {
3
+ "choice:3": 1.0175427143279028,
4
+ "choice:4": 1.3738458063981829
5
+ },
6
+ "source": "historical v3 calibration split; 3923 samples",
7
+ "group_counts": {
8
+ "choice:3": 1887,
9
+ "choice:4": 2036
10
+ },
11
+ "prompt_version": "audio-jev-labels-v4"
12
+ }
image/01_metric_comparison.png ADDED

Git LFS Details

  • SHA256: 797025024168f1c8e6cd26e0b8b4c3286ef1f1214bbe7e7efe57a781bbbfec12
  • Pointer size: 131 Bytes
  • Size of remote file: 404 kB
image/02_latency_comparison.png ADDED

Git LFS Details

  • SHA256: 1beeaa8bde006079f467590a6f014c68498ebfad67eb678b3734294db2e41b46
  • Pointer size: 131 Bytes
  • Size of remote file: 429 kB
image/效果对比.png ADDED

Git LFS Details

  • SHA256: 797025024168f1c8e6cd26e0b8b4c3286ef1f1214bbe7e7efe57a781bbbfec12
  • Pointer size: 131 Bytes
  • Size of remote file: 404 kB
release_manifest.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "complete_model": false,
3
+ "weight_files": [],
4
+ "links_are_placeholders": true,
5
+ "hf_repo_id": "YOUR_HF_ORG/Turn-Detection-Jev",
6
+ "github_repo": "YOUR_GITHUB_ORG/Turn-Detection-Jev",
7
+ "github_ref": "main",
8
+ "sha256": {
9
+ ".gitattributes": "77db011f87feff0d3ed5a3d8935c8df0ccdf1f3f0e37ad39de66ef56aa3dc130",
10
+ "EVALUATION.md": "12cbe3a42811b407d6589b72d635dfd2be1ed0926e3401f1c75b515257b91905",
11
+ "LICENSE": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
12
+ "NOTICE": "0fa613f92457ef19139f198d3ea35e9f2879ea552f886ddf82977bd485102d62",
13
+ "README.md": "5d2e3912da2d40ba9f9abf37e49aee159fa8ed1f671635837ba437943c2856ae",
14
+ "README_zh.md": "736c25f1530929eb8b048b88eb7972ac98b09da71c0a91738427ada9e286fdd4",
15
+ "calibration.json": "e0bdf32ab8aae2d0750b9d50ace497cf5cd2d1729ad346063f3ec13996eb3a3e",
16
+ "image/01_metric_comparison.png": "797025024168f1c8e6cd26e0b8b4c3286ef1f1214bbe7e7efe57a781bbbfec12",
17
+ "image/02_latency_comparison.png": "1beeaa8bde006079f467590a6f014c68498ebfad67eb678b3734294db2e41b46",
18
+ "image/效果对比.png": "797025024168f1c8e6cd26e0b8b4c3286ef1f1214bbe7e7efe57a781bbbfec12"
19
+ },
20
+ "gpu_validation": "not performed by packaging; actual checkpoint validation remains required"
21
+ }