Hugging Face's logo

AXERA-TECH
/
FS-EEND.AXERA

Audio Classification
axera
speaker-diarization
ls-eend
fs-eend
end-to-end-neural-diarization
rttm
speech
ax650n
Model card Files
xet
 Community
FS-EEND.AXERA
File size: 2,449 Bytes
3d059cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
{
  "model_name": "ls_eend_streaming_step",
  "task": "streaming_speaker_diarization",
  "input_names": [
    "feat",
    "inv_count",
    "dec_inv_count",
    "conv_cache",
    "enc0_kv",
    "enc0_conv",
    "enc1_kv",
    "enc1_conv",
    "enc2_kv",
    "enc2_conv",
    "enc3_kv",
    "enc3_conv",
    "dec0_kv",
    "dec1_kv"
  ],
  "output_names": [
    "pred",
    "enc0_inc",
    "enc0_conv_out",
    "enc1_inc",
    "enc1_conv_out",
    "enc2_inc",
    "enc2_conv_out",
    "enc3_inc",
    "enc3_conv_out",
    "conv_cache_out",
    "dec0_inc",
    "dec1_inc"
  ],
  "input_shapes": {
    "feat": [
      1,
      1,
      345
    ],
    "inv_count": [
      1,
      4,
      1,
      1
    ],
    "dec_inv_count": [
      1,
      4,
      1,
      1
    ],
    "conv_cache": [
      1,
      256,
      18
    ],
    "enc0_kv": [
      1,
      4,
      64,
      64
    ],
    "enc0_conv": [
      1,
      256,
      15
    ],
    "enc1_kv": [
      1,
      4,
      64,
      64
    ],
    "enc1_conv": [
      1,
      256,
      15
    ],
    "enc2_kv": [
      1,
      4,
      64,
      64
    ],
    "enc2_conv": [
      1,
      256,
      15
    ],
    "enc3_kv": [
      1,
      4,
      64,
      64
    ],
    "enc3_conv": [
      1,
      256,
      15
    ],
    "dec0_kv": [
      10,
      4,
      64,
      64
    ],
    "dec1_kv": [
      10,
      4,
      64,
      64
    ]
  },
  "notes": "frame-wise state model in bounded mean form. Retention states hold the running mean K_t/t instead of the upstream K_t/sqrt(t), so magnitudes stay bounded and a single U16 scale covers the whole recording; the dropped sqrt(t) gain is cancelled exactly by the following affine-free LayerNorm. The graph emits the bounded per-frame increment enc{i}_inc/dec{i}_inc, and the CALLER accumulates mean += (inc-mean)/t in FP32 so quantization error stays per-frame. Caller also feeds inv_count=1/t (encoder) and dec_inv_count=1/t_dec (decoder), and MUST hold decoder state frozen for the first 9 frames (conv warmup) while discarding those preds, matching native StreamingConv1d.",
  "torch_onnx_one_step_min_cosine": 0.9999998807907104,
  "torch_onnx_one_step_max_diff": 0.0009765625,
  "torch_onnx_one_step_output_cosines": [
    1.0000001192092896,
    1.0,
    1.0000001192092896,
    1.0000001192092896,
    1.0,
    1.0,
    0.9999999403953552,
    1.0,
    0.9999998807907104,
    1.0,
    1.0,
    1.000000238418579
  ]
}