Skip to content

vap_prompt API

maai.models.vap_prompt

VapGPT_prompt

Bases: Module

Voice Activity Projection with Prompt Control (Beta).

This model integrates text prompts (e.g., personality or instruction prompts) into the VAP architecture using sentence embeddings to condition the output.

Source code in src/maai/models/vap_prompt.py 
 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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
class VapGPT_prompt(nn.Module):
    """Voice Activity Projection with Prompt Control (Beta).

    This model integrates text prompts (e.g., personality or instruction prompts)
    into the VAP architecture using sentence embeddings to condition the output.
    """
    BINS_P_NOW = [0, 1]
    BINS_PFUTURE = [2, 3]

    # prompt_model_name = "sbintuitions/sarashina-embedding-v1-1b"
    prompt_model_name = "cl-nagoya/ruri-v3-pt-30m"

    def __init__(self, conf: Optional[VapConfig] = None):
        """Initialize the VapGPT_prompt model.

        Args:
            conf (Optional[VapConfig]): Configuration object.
                If None, default VapConfig is used.
        """

        # print this model is a beta version
        print('--------------------------------')
        print("<<< This is a beta version of model !!! >>>")
        print("VAP with prompt control is under development.")
        print("This is a beta version of VapGPT with prompt support. It may not work as expected.")
        print("This model also requires 'sentence-transformers protobuf sentencepiece' package to be installed.")
        print('--------------------------------')

        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.ar_channel = 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.ar = 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)
        self.vap_head = nn.Linear(conf.dim, self.objective.n_classes)
        self.prompt_head = nn.Linear(conf.dim, conf.dim_prompt)

        self.prompt_embed1 = nn.Linear(self.conf.dim_prompt, self.conf.dim_prompt_2)
        self.prompt_embed2 = nn.Linear(self.conf.dim_prompt, self.conf.dim_prompt_2)

        self.prompt_dim_red1 = nn.Linear(self.conf.dim + self.conf.dim_prompt_2, self.conf.dim)
        self.prompt_dim_red2 = nn.Linear(self.conf.dim + self.conf.dim_prompt_2, self.conf.dim)

        # Initialize the embedding model for prompts
        print("Loading prompt embedding model:", self.prompt_model_name)
        self.prompt_embedding_model = SentenceTransformer(self.prompt_model_name)
        print("Prompt embedding model loaded.")

        # Initialize the prompt embeddings
        self.set_prompt_ch1("テンポよく発話し、相手の発言が終わるとすぐに返答してください。発言回数を多めに、会話をリードするようにしてください。")
        self.set_prompt_ch2("発話前に少し間を取り、考えてから丁寧に話し始めてください。応答は急がず、落ち着いたテンポを意識してください。")

    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.

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

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

        x1 = self.encoder1(audio1)  # speaker 1
        x2 = self.encoder2(audio2)  # speaker 2

        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 set_prompt_ch1(self, prompt: str, device: torch.device = torch.device('cpu')):
        """Set the text prompt for channel 1 (speaker 1).

        Args:
            prompt (str): The text instruction or personality prompt.
            device (torch.device): The device to load the embedding tensor onto.
        """

        embedding_ch1_ = self.prompt_embedding_model.encode([prompt], normalize_embeddings=True)[0]
        self.embedding_ch1 = torch.tensor(embedding_ch1_).unsqueeze(0)
        self.embedding_ch1 = self.embedding_ch1.to(device=device, non_blocking=True)

        # print("Embedding for channel 1 set:", self.embedding_ch1.shape)
        # print("Embedding for channel 1:", self.embedding_ch1)
        # input("Press Enter to continue...")

    def set_prompt_ch2(self, prompt: str, device: torch.device = torch.device('cpu')):
        """Set the text prompt for channel 2 (speaker 2).

        Args:
            prompt (str): The text instruction or personality prompt.
            device (torch.device): The device to load the embedding tensor onto.
        """

        embedding_ch2_ = self.prompt_embedding_model.encode([prompt], normalize_embeddings=True)[0]
        self.embedding_ch2 = torch.tensor(embedding_ch2_).unsqueeze(0)
        self.embedding_ch2 = self.embedding_ch2.to(device=device, non_blocking=True)

        # print("Embedding for channel 2 set:", self.embedding_ch2.shape)
        # print("Embedding for channel 2:", self.embedding_ch2)
        # input("Press Enter to continue...")

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

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

        Returns:
            Tuple[dict, dict]: Output tensors and updated cache.
        """

        if cache is None:
            cache = {}

        prompt_a_seq = self.embedding_ch1.unsqueeze(1).repeat(1, x1.shape[1], 1)
        prompt_b_seq = self.embedding_ch2.unsqueeze(1).repeat(1, x2.shape[1], 1)

        # print("prompt_a_seq", prompt_a_seq.shape)
        # print("prompt_b_seq", prompt_b_seq.shape)
        prompt_a_embed1 = self.prompt_embed1(prompt_a_seq)
        prompt_b_embed1 = self.prompt_embed1(prompt_b_seq)

        # print("prompt_a_embed1", prompt_a_embed1.shape)
        # print("prompt_b_embed1", prompt_b_embed1.shape)

        # print("x1", x1.shape)
        # print("x2", x2.shape)

        x1_concat = torch.cat((x1, prompt_a_embed1), dim=-1)
        x2_concat = torch.cat((x2, prompt_b_embed1), dim=-1)
        x1_concat = self.prompt_dim_red1(x1_concat)
        x2_concat = self.prompt_dim_red1(x2_concat)

        o1 = self.ar_channel(x1_concat, past_kv=cache.get("ar1"))
        o2 = self.ar_channel(x2_concat, past_kv=cache.get("ar2"))

        # o1_red = self.prompt_dim_red1(o1["x"])
        # o2_red = self.prompt_dim_red1(o2["x"])

        prompt_a_embed2 = self.prompt_embed2(prompt_a_seq)
        prompt_b_embed2 = self.prompt_embed2(prompt_b_seq)

        o1_concat = torch.cat((o1["x"], prompt_a_embed2), dim=-1)
        o2_concat = torch.cat((o2["x"], prompt_b_embed2), dim=-1)

        o1_concat = self.prompt_dim_red2(o1_concat)
        o2_concat = self.prompt_dim_red2(o2_concat)

        out = self.ar(
            o1_concat,
            o2_concat,
            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"]),
        }

        # Outputs
        vad1 = self.va_classifier(out["x1"])
        vad2 = self.va_classifier(out["x2"])
        logits = self.vap_head(out["x"])

        # print("logits", logits.shape)
        # print(logits)
        probs = logits.softmax(dim=-1)

        p_now = self.objective.probs_next_speaker_aggregate(
            probs,
            from_bin=self.BINS_P_NOW[0],
            to_bin=self.BINS_P_NOW[-1]
        )

        p_future = self.objective.probs_next_speaker_aggregate(
            probs,
            from_bin=self.BINS_PFUTURE[0],
            to_bin=self.BINS_PFUTURE[1]
        )

        # Get back to the CPU
        p_now = p_now.to('cpu').tolist()[0][-1]
        p_future = p_future.to('cpu').tolist()[0][-1]

        vad1 = vad1.sigmoid().to('cpu').tolist()[0][-1][0]
        vad2 = vad2.sigmoid().to('cpu').tolist()[0][-1][0]

        ret = {"p_now": p_now, "p_future": p_future, "vad": [vad1, vad2]}

        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_prompt 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_prompt.py 
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
def __init__(self, conf: Optional[VapConfig] = None):
    """Initialize the VapGPT_prompt model.

    Args:
        conf (Optional[VapConfig]): Configuration object.
            If None, default VapConfig is used.
    """

    # print this model is a beta version
    print('--------------------------------')
    print("<<< This is a beta version of model !!! >>>")
    print("VAP with prompt control is under development.")
    print("This is a beta version of VapGPT with prompt support. It may not work as expected.")
    print("This model also requires 'sentence-transformers protobuf sentencepiece' package to be installed.")
    print('--------------------------------')

    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.ar_channel = 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.ar = 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)
    self.vap_head = nn.Linear(conf.dim, self.objective.n_classes)
    self.prompt_head = nn.Linear(conf.dim, conf.dim_prompt)

    self.prompt_embed1 = nn.Linear(self.conf.dim_prompt, self.conf.dim_prompt_2)
    self.prompt_embed2 = nn.Linear(self.conf.dim_prompt, self.conf.dim_prompt_2)

    self.prompt_dim_red1 = nn.Linear(self.conf.dim + self.conf.dim_prompt_2, self.conf.dim)
    self.prompt_dim_red2 = nn.Linear(self.conf.dim + self.conf.dim_prompt_2, self.conf.dim)

    # Initialize the embedding model for prompts
    print("Loading prompt embedding model:", self.prompt_model_name)
    self.prompt_embedding_model = SentenceTransformer(self.prompt_model_name)
    print("Prompt embedding model loaded.")

    # Initialize the prompt embeddings
    self.set_prompt_ch1("テンポよく発話し、相手の発言が終わるとすぐに返答してください。発言回数を多めに、会話をリードするようにしてください。")
    self.set_prompt_ch2("発話前に少し間を取り、考えてから丁寧に話し始めてください。応答は急がず、落ち着いたテンポを意識してください。")

encode_audio(audio1, audio2)

Encode the raw audio inputs into feature representations.

Parameters:

Name Type Description Default
audio1 Tensor

Audio waveform for speaker 1.

 required
audio2 Tensor

Audio waveform for speaker 2.

 required

Returns:

Type Description
Tuple[Tensor, Tensor]

Tuple[Tensor, Tensor]: Encoded features for the two speakers.
Source code in src/maai/models/vap_prompt.py 
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def encode_audio(self, audio1: torch.Tensor, audio2: torch.Tensor) -> Tuple[Tensor, Tensor]:
    """Encode the raw audio inputs into feature representations.

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

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

    x1 = self.encoder1(audio1)  # speaker 1
    x2 = self.encoder2(audio2)  # speaker 2

    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_prompt model.

Parameters:

Name Type Description Default
x1 Tensor

Input audio tensor for speaker 1.

 required
x2 Tensor

Input audio tensor for speaker 2.

 required
cache dict

Cache of past keys/values.

 None

Returns:

Type Description
Tuple[dict, dict]

Tuple[dict, dict]: Output tensors and updated cache.
Source code in src/maai/models/vap_prompt.py 
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
def forward(
    self,
    x1: Tensor,
    x2: Tensor,
    cache: Optional[dict] = None,
) -> Tuple[dict, dict]:
    """
    Forward pass for the VapGPT_prompt model.

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

    Returns:
        Tuple[dict, dict]: Output tensors and updated cache.
    """

    if cache is None:
        cache = {}

    prompt_a_seq = self.embedding_ch1.unsqueeze(1).repeat(1, x1.shape[1], 1)
    prompt_b_seq = self.embedding_ch2.unsqueeze(1).repeat(1, x2.shape[1], 1)

    # print("prompt_a_seq", prompt_a_seq.shape)
    # print("prompt_b_seq", prompt_b_seq.shape)
    prompt_a_embed1 = self.prompt_embed1(prompt_a_seq)
    prompt_b_embed1 = self.prompt_embed1(prompt_b_seq)

    # print("prompt_a_embed1", prompt_a_embed1.shape)
    # print("prompt_b_embed1", prompt_b_embed1.shape)

    # print("x1", x1.shape)
    # print("x2", x2.shape)

    x1_concat = torch.cat((x1, prompt_a_embed1), dim=-1)
    x2_concat = torch.cat((x2, prompt_b_embed1), dim=-1)
    x1_concat = self.prompt_dim_red1(x1_concat)
    x2_concat = self.prompt_dim_red1(x2_concat)

    o1 = self.ar_channel(x1_concat, past_kv=cache.get("ar1"))
    o2 = self.ar_channel(x2_concat, past_kv=cache.get("ar2"))

    # o1_red = self.prompt_dim_red1(o1["x"])
    # o2_red = self.prompt_dim_red1(o2["x"])

    prompt_a_embed2 = self.prompt_embed2(prompt_a_seq)
    prompt_b_embed2 = self.prompt_embed2(prompt_b_seq)

    o1_concat = torch.cat((o1["x"], prompt_a_embed2), dim=-1)
    o2_concat = torch.cat((o2["x"], prompt_b_embed2), dim=-1)

    o1_concat = self.prompt_dim_red2(o1_concat)
    o2_concat = self.prompt_dim_red2(o2_concat)

    out = self.ar(
        o1_concat,
        o2_concat,
        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"]),
    }

    # Outputs
    vad1 = self.va_classifier(out["x1"])
    vad2 = self.va_classifier(out["x2"])
    logits = self.vap_head(out["x"])

    # print("logits", logits.shape)
    # print(logits)
    probs = logits.softmax(dim=-1)

    p_now = self.objective.probs_next_speaker_aggregate(
        probs,
        from_bin=self.BINS_P_NOW[0],
        to_bin=self.BINS_P_NOW[-1]
    )

    p_future = self.objective.probs_next_speaker_aggregate(
        probs,
        from_bin=self.BINS_PFUTURE[0],
        to_bin=self.BINS_PFUTURE[1]
    )

    # Get back to the CPU
    p_now = p_now.to('cpu').tolist()[0][-1]
    p_future = p_future.to('cpu').tolist()[0][-1]

    vad1 = vad1.sigmoid().to('cpu').tolist()[0][-1][0]
    vad2 = vad2.sigmoid().to('cpu').tolist()[0][-1][0]

    ret = {"p_now": p_now, "p_future": p_future, "vad": [vad1, vad2]}

    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_prompt.py 
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
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()

set_prompt_ch1(prompt, device=torch.device('cpu'))

Set the text prompt for channel 1 (speaker 1).

Parameters:

Name Type Description Default
prompt str

The text instruction or personality prompt.

 required
device device

The device to load the embedding tensor onto.

 device('cpu')
Source code in src/maai/models/vap_prompt.py 
154
155
156
157
158
159
160
161
162
163
164
def set_prompt_ch1(self, prompt: str, device: torch.device = torch.device('cpu')):
    """Set the text prompt for channel 1 (speaker 1).

    Args:
        prompt (str): The text instruction or personality prompt.
        device (torch.device): The device to load the embedding tensor onto.
    """

    embedding_ch1_ = self.prompt_embedding_model.encode([prompt], normalize_embeddings=True)[0]
    self.embedding_ch1 = torch.tensor(embedding_ch1_).unsqueeze(0)
    self.embedding_ch1 = self.embedding_ch1.to(device=device, non_blocking=True)

set_prompt_ch2(prompt, device=torch.device('cpu'))

Set the text prompt for channel 2 (speaker 2).

Parameters:

Name Type Description Default
prompt str

The text instruction or personality prompt.

 required
device device

The device to load the embedding tensor onto.

 device('cpu')
Source code in src/maai/models/vap_prompt.py 
170
171
172
173
174
175
176
177
178
179
180
def set_prompt_ch2(self, prompt: str, device: torch.device = torch.device('cpu')):
    """Set the text prompt for channel 2 (speaker 2).

    Args:
        prompt (str): The text instruction or personality prompt.
        device (torch.device): The device to load the embedding tensor onto.
    """

    embedding_ch2_ = self.prompt_embedding_model.encode([prompt], normalize_embeddings=True)[0]
    self.embedding_ch2 = torch.tensor(embedding_ch2_).unsqueeze(0)
    self.embedding_ch2 = self.embedding_ch2.to(device=device, non_blocking=True)

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_prompt.py 
142
143
144
145
146
147
148
149
150
151
152
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)