Skip to content

vap_nod API

maai.models.vap_nod

VapGPT_nod

Bases: Module

Voice Activity Projection with Nodding prediction (VAP-Nod) model.

This model predicts both backchannels and nodding behaviors (e.g., short nod, long nod) during a conversation.

Source code in src/maai/models/vap_nod.py 
 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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
class VapGPT_nod(nn.Module):
    """Voice Activity Projection with Nodding prediction (VAP-Nod) model.

    This model predicts both backchannels and nodding behaviors (e.g., short nod,
    long nod) during a conversation.
    """
    def __init__(self, conf: Optional[VapConfig] = None):
        """Initialize the VapGPT_nod model.

        Args:
            conf (Optional[VapConfig]): Configuration object.
                If None, default VapConfig is used.
        """
        super().__init__()
        if conf is None:
            conf = VapConfig()
        self.conf = conf
        self.sample_rate = conf.sample_rate
        self.frame_hz = conf.frame_hz

        self.temp_elapse_time = []

        # Single channel
        self.self_attention = GPT(
            dim=conf.dim,
            dff_k=3,
            num_layers=conf.channel_layers,
            num_heads=conf.num_heads,
            dropout=conf.dropout,
            context_limit=conf.context_limit,
        )

        # Cross channel
        self.cross_attention = GPTStereo(
            dim=conf.dim,
            dff_k=3,
            num_layers=conf.cross_layers,
            num_heads=conf.num_heads,
            dropout=conf.dropout,
            context_limit=conf.context_limit,
        )

        self.objective = ObjectiveVAP(bin_times=conf.bin_times, frame_hz=conf.frame_hz)

        # Outputs
        # Voice activity objective -> x1, x2 -> logits ->  BCE
        self.va_classifier = nn.Linear(conf.dim, 1)

        if self.conf.lid_classify == 1:
            self.lid_classifier = nn.Linear(conf.dim, conf.lid_classify_num_class)

        elif self.conf.lid_classify == 2:
            self.lid_classifier_middle = nn.Linear(conf.dim*2, conf.lid_classify_num_class)

        if self.conf.lang_cond == 1:
            self.lang_condition = nn.Linear(conf.lid_classify_num_class, conf.dim)

        self.vap_head = nn.Linear(conf.dim, self.objective.n_classes)

        # For Nodding
        self.gt_head = nn.Linear(conf.dim, 4)
        self.bc_head = nn.Linear(conf.dim, 1)

    def load_encoder(self, cpc_model):
        """Load and build the audio encoders for both speakers.

        Args:
            cpc_model: Pre-trained CPC model to be used as feature extractor.
        """
        self.encoder1 = build_audio_encoder(self.conf, cpc_model=cpc_model)
        self.encoder1 = self.encoder1.eval()
        self.encoder2 = build_audio_encoder(self.conf, cpc_model=cpc_model)
        self.encoder2 = self.encoder2.eval()

        encoder_dim = getattr(self.encoder1, "output_dim", self.conf.dim)
        if encoder_dim != self.conf.dim:
            self.decrease_dimension = nn.Linear(encoder_dim, self.conf.dim)

        if self.conf.freeze_encoder == 1:
            print('freeze encoder')
            self.encoder1.freeze()
            self.encoder2.freeze()

    @property
    def horizon_time(self):
        """Get the horizon time for the projection in seconds.

        Returns:
            float: Horizon time for the objective.
        """
        return self.objective.horizon_time

    def encode_audio(self, audio1: torch.Tensor, audio2: torch.Tensor) -> Tuple[Tensor, Tensor]:
        """Encode the raw audio inputs into feature representations.

        Note: Channel swap is applied for temporal consistency.

        Args:
            audio1 (torch.Tensor): Audio waveform for speaker 1 (User).
            audio2 (torch.Tensor): Audio waveform for speaker 2 (System).

        Returns:
            Tuple[Tensor, Tensor]: Encoded features for the two speakers.
        """

        # Channel swap for temporal consistency
        x1 = self.encoder1(audio2)  # speaker 1 (User)
        x2 = self.encoder2(audio1)  # speaker 2 (System)

        if hasattr(self, "decrease_dimension"):
            x1 = self.decrease_dimension(x1)
            x2 = self.decrease_dimension(x2)

        return x1, x2

    def vad_loss(self, vad_output, vad):
        """Compute the Voice Activity Detection (VAD) loss.

        Args:
            vad_output: Predicted VAD logits.
            vad: Ground truth VAD labels.

        Returns:
            Tensor: Binary cross-entropy loss between predictions and targets.
        """
        return F.binary_cross_entropy_with_logits(vad_output, vad)

    def forward(
        self,
        x1: Tensor,
        x2: Tensor,
        cache: Optional[dict] = None,
    ) -> Tuple[dict, dict]:
        """
        Forward pass for the VapGPT_nod model.

        Args:
            x1 (Tensor): Input audio embedded tensor for speaker 1.
            x2 (Tensor): Input audio embedded tensor for speaker 2.
            cache (dict, optional): Cache of past keys/values.

        Returns:
            Tuple[dict, dict]: Model outputs and updated cache.
        """

        if cache is None:
            cache = {}

        o1 = self.self_attention(x1, past_kv=cache.get("ar1"))
        o2 = self.self_attention(x2, past_kv=cache.get("ar2"))
        out = self.cross_attention(
            o1["x"],
            o2["x"],
            past_kv1=cache.get("cross1"),
            past_kv2=cache.get("cross2"),
            past_kv1_c=cache.get("cross1_c"),
            past_kv2_c=cache.get("cross2_c"),
        )

        new_cache = {
            "ar1": (o1["past_k"], o1["past_v"]),
            "ar2": (o2["past_k"], o2["past_v"]),
            "cross1": (out["past_k1"], out["past_v1"]),
            "cross2": (out["past_k2"], out["past_v2"]),
            "cross1_c": (out["past_k1_c"], out["past_v1_c"]),
            "cross2_c": (out["past_k2_c"], out["past_v2_c"]),
        }

        p_bc = self.bc_head(out["x"])
        nod = self.gt_head(out["x"])

        p_bc = p_bc.sigmoid().to("cpu").tolist()[0][-1][0]
        nod_ = nod.softmax(dim=-1).to("cpu").tolist()[0][-1]
        p_nod_short = nod_[1]
        p_nod_long = nod_[2]
        p_nod_long_p = nod_[3]

        ret = {
            "p_bc": p_bc,
            "p_nod_short": p_nod_short,
            "p_nod_long": p_nod_long,
            "p_nod_long_p": p_nod_long_p,
        }

        return ret, new_cache

horizon_time property

Get the horizon time for the projection in seconds.

Returns:

Name Type Description
float

Horizon time for the objective.

__init__(conf=None)

Initialize the VapGPT_nod model.

Parameters:

Name Type Description Default
conf Optional[VapConfig]

Configuration object. If None, default VapConfig is used.

 None
Source code in src/maai/models/vap_nod.py 
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
def __init__(self, conf: Optional[VapConfig] = None):
    """Initialize the VapGPT_nod model.

    Args:
        conf (Optional[VapConfig]): Configuration object.
            If None, default VapConfig is used.
    """
    super().__init__()
    if conf is None:
        conf = VapConfig()
    self.conf = conf
    self.sample_rate = conf.sample_rate
    self.frame_hz = conf.frame_hz

    self.temp_elapse_time = []

    # Single channel
    self.self_attention = GPT(
        dim=conf.dim,
        dff_k=3,
        num_layers=conf.channel_layers,
        num_heads=conf.num_heads,
        dropout=conf.dropout,
        context_limit=conf.context_limit,
    )

    # Cross channel
    self.cross_attention = GPTStereo(
        dim=conf.dim,
        dff_k=3,
        num_layers=conf.cross_layers,
        num_heads=conf.num_heads,
        dropout=conf.dropout,
        context_limit=conf.context_limit,
    )

    self.objective = ObjectiveVAP(bin_times=conf.bin_times, frame_hz=conf.frame_hz)

    # Outputs
    # Voice activity objective -> x1, x2 -> logits ->  BCE
    self.va_classifier = nn.Linear(conf.dim, 1)

    if self.conf.lid_classify == 1:
        self.lid_classifier = nn.Linear(conf.dim, conf.lid_classify_num_class)

    elif self.conf.lid_classify == 2:
        self.lid_classifier_middle = nn.Linear(conf.dim*2, conf.lid_classify_num_class)

    if self.conf.lang_cond == 1:
        self.lang_condition = nn.Linear(conf.lid_classify_num_class, conf.dim)

    self.vap_head = nn.Linear(conf.dim, self.objective.n_classes)

    # For Nodding
    self.gt_head = nn.Linear(conf.dim, 4)
    self.bc_head = nn.Linear(conf.dim, 1)

encode_audio(audio1, audio2)

Encode the raw audio inputs into feature representations.

Note: Channel swap is applied for temporal consistency.

Parameters:

Name Type Description Default
audio1 Tensor

Audio waveform for speaker 1 (User).

 required
audio2 Tensor

Audio waveform for speaker 2 (System).

 required

Returns:

Type Description
Tuple[Tensor, Tensor]

Tuple[Tensor, Tensor]: Encoded features for the two speakers.
Source code in src/maai/models/vap_nod.py 
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
def encode_audio(self, audio1: torch.Tensor, audio2: torch.Tensor) -> Tuple[Tensor, Tensor]:
    """Encode the raw audio inputs into feature representations.

    Note: Channel swap is applied for temporal consistency.

    Args:
        audio1 (torch.Tensor): Audio waveform for speaker 1 (User).
        audio2 (torch.Tensor): Audio waveform for speaker 2 (System).

    Returns:
        Tuple[Tensor, Tensor]: Encoded features for the two speakers.
    """

    # Channel swap for temporal consistency
    x1 = self.encoder1(audio2)  # speaker 1 (User)
    x2 = self.encoder2(audio1)  # speaker 2 (System)

    if hasattr(self, "decrease_dimension"):
        x1 = self.decrease_dimension(x1)
        x2 = self.decrease_dimension(x2)

    return x1, x2

forward(x1, x2, cache=None)

Forward pass for the VapGPT_nod model.

Parameters:

Name Type Description Default
x1 Tensor

Input audio embedded tensor for speaker 1.

 required
x2 Tensor

Input audio embedded tensor for speaker 2.

 required
cache dict

Cache of past keys/values.

 None

Returns:

Type Description
Tuple[dict, dict]

Tuple[dict, dict]: Model outputs and updated cache.
Source code in src/maai/models/vap_nod.py 
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
def forward(
    self,
    x1: Tensor,
    x2: Tensor,
    cache: Optional[dict] = None,
) -> Tuple[dict, dict]:
    """
    Forward pass for the VapGPT_nod model.

    Args:
        x1 (Tensor): Input audio embedded tensor for speaker 1.
        x2 (Tensor): Input audio embedded tensor for speaker 2.
        cache (dict, optional): Cache of past keys/values.

    Returns:
        Tuple[dict, dict]: Model outputs and updated cache.
    """

    if cache is None:
        cache = {}

    o1 = self.self_attention(x1, past_kv=cache.get("ar1"))
    o2 = self.self_attention(x2, past_kv=cache.get("ar2"))
    out = self.cross_attention(
        o1["x"],
        o2["x"],
        past_kv1=cache.get("cross1"),
        past_kv2=cache.get("cross2"),
        past_kv1_c=cache.get("cross1_c"),
        past_kv2_c=cache.get("cross2_c"),
    )

    new_cache = {
        "ar1": (o1["past_k"], o1["past_v"]),
        "ar2": (o2["past_k"], o2["past_v"]),
        "cross1": (out["past_k1"], out["past_v1"]),
        "cross2": (out["past_k2"], out["past_v2"]),
        "cross1_c": (out["past_k1_c"], out["past_v1_c"]),
        "cross2_c": (out["past_k2_c"], out["past_v2_c"]),
    }

    p_bc = self.bc_head(out["x"])
    nod = self.gt_head(out["x"])

    p_bc = p_bc.sigmoid().to("cpu").tolist()[0][-1][0]
    nod_ = nod.softmax(dim=-1).to("cpu").tolist()[0][-1]
    p_nod_short = nod_[1]
    p_nod_long = nod_[2]
    p_nod_long_p = nod_[3]

    ret = {
        "p_bc": p_bc,
        "p_nod_short": p_nod_short,
        "p_nod_long": p_nod_long,
        "p_nod_long_p": p_nod_long_p,
    }

    return ret, new_cache

load_encoder(cpc_model)

Load and build the audio encoders for both speakers.

Parameters:

Name Type Description Default
cpc_model

Pre-trained CPC model to be used as feature extractor.

 required
Source code in src/maai/models/vap_nod.py 
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
def load_encoder(self, cpc_model):
    """Load and build the audio encoders for both speakers.

    Args:
        cpc_model: Pre-trained CPC model to be used as feature extractor.
    """
    self.encoder1 = build_audio_encoder(self.conf, cpc_model=cpc_model)
    self.encoder1 = self.encoder1.eval()
    self.encoder2 = build_audio_encoder(self.conf, cpc_model=cpc_model)
    self.encoder2 = self.encoder2.eval()

    encoder_dim = getattr(self.encoder1, "output_dim", self.conf.dim)
    if encoder_dim != self.conf.dim:
        self.decrease_dimension = nn.Linear(encoder_dim, self.conf.dim)

    if self.conf.freeze_encoder == 1:
        print('freeze encoder')
        self.encoder1.freeze()
        self.encoder2.freeze()

vad_loss(vad_output, vad)

Compute the Voice Activity Detection (VAD) loss.

Parameters:

Name Type Description Default
vad_output

Predicted VAD logits.

 required
vad

Ground truth VAD labels.

 required

Returns:

Name Type Description
Tensor

Binary cross-entropy loss between predictions and targets.
Source code in src/maai/models/vap_nod.py 
127
128
129
130
131
132
133
134
135
136
137
def vad_loss(self, vad_output, vad):
    """Compute the Voice Activity Detection (VAD) loss.

    Args:
        vad_output: Predicted VAD logits.
        vad: Ground truth VAD labels.

    Returns:
        Tensor: Binary cross-entropy loss between predictions and targets.
    """
    return F.binary_cross_entropy_with_logits(vad_output, vad)