Skip to content

vap_mono API

maai.models.vap_mono

VapGPT_mono

Bases: Module

Single-channel (mono) Voice Activity Projection (VAP) model.

Unlike the two-speaker :class:~maai.models.vap.VapGPT, this is a genuinely single-channel model trained on its own vap_mono checkpoints: it encodes one audio stream, replaces the cross-channel GPTStereo with a plain causal :class:~maai.modules.GPT, and predicts 2 ** n_bins (16 for the default four bins) codebook states describing the future activity of that one speaker.

Because there is no second speaker to normalize against, p_now and p_future are not a two-way split: they are the expected (bin-length weighted) activity ratio of the input speaker and are already in [0, 1]. They are returned as single floats, as are vad, p_bins_now and p_bins_future; p_bins is a list with one value per bin.

It is intended for use cases where only one speaker's audio is available, e.g. a single microphone input for a spoken dialogue system. Maai uses it for mode="vap" when audio_ch2 is omitted (or is a MaaiInput.Zero).

The same class also serves the vap_mc_mono checkpoints (selected with mc=True), which are trained with noise and random-gain augmentation (multi-condition) for robustness in real-world environments.

Source code in src/maai/models/vap_mono.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
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
class VapGPT_mono(nn.Module):
    """Single-channel (mono) Voice Activity Projection (VAP) model.

    Unlike the two-speaker :class:`~maai.models.vap.VapGPT`, this is a genuinely
    single-channel model trained on its own ``vap_mono`` checkpoints: it encodes
    one audio stream, replaces the cross-channel ``GPTStereo`` with a plain
    causal :class:`~maai.modules.GPT`, and predicts ``2 ** n_bins`` (16 for the
    default four bins) codebook states describing the future activity of that
    one speaker.

    Because there is no second speaker to normalize against, ``p_now`` and
    ``p_future`` are *not* a two-way split: they are the expected (bin-length
    weighted) activity ratio of the input speaker and are already in [0, 1].
    They are returned as single floats, as are ``vad``, ``p_bins_now`` and
    ``p_bins_future``; ``p_bins`` is a list with one value per bin.

    It is intended for use cases where only one speaker's audio is available,
    e.g. a single microphone input for a spoken dialogue system. ``Maai`` uses
    it for ``mode="vap"`` when ``audio_ch2`` is omitted (or is a
    ``MaaiInput.Zero``).

    The same class also serves the ``vap_mc_mono`` checkpoints (selected with
    ``mc=True``), which are trained with noise and random-gain augmentation
    (multi-condition) for robustness in real-world environments.
    """

    # Only channel 1 is encoded, so the runtime must not feed a second stream
    # through ``encode_audio`` / ``forward``.
    is_single_tower = True

    BINS_P_NOW = [0, 1]
    BINS_PFUTURE = [2, 3]

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

        Args:
            conf (Optional[VapConfig]): Configuration object for the model.
                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.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,
        )

        # There is no second channel to cross-attend to, so the stereo tower of
        # the two-speaker model is a plain causal GPT here. The layer counts are
        # kept so the depth matches the two-speaker model.
        self.ar = GPT(
            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, num_channels=1
        )

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

    def load_encoder(self, cpc_model):
        """Load and build the audio encoder for the single input channel.

        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()

        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()

    @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) -> Tensor:
        """Encode the raw audio input into a feature representation.

        Args:
            audio1 (torch.Tensor): Audio waveform for the input speaker.

        Returns:
            Tensor: Encoded features for the input speaker.
        """
        x1 = self.encoder1(audio1)

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

        return x1

    def forward(
        self,
        x1: Tensor,
        cache: Optional[dict] = None,
        return_all_frames: bool = False,
    ) -> Tuple[dict, dict]:
        """
        Forward pass for the mono VapGPT model.

        Args:
            x1 (Tensor): Input audio embedded tensor for the input speaker.
            cache (dict, optional): Cache of past keys/values.
            return_all_frames: Return one result dictionary for every input
                time step instead of only the final step.

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

        if cache is None:
            cache = {}

        o1 = self.ar_channel(x1, past_kv=cache.get("ar1"))
        out = self.ar(o1["x"], past_kv=cache.get("ar2"))

        new_cache = {
            "ar1": (o1["past_k"], o1["past_v"]),
            "ar2": (out["past_k"], out["past_v"]),
        }

        # Outputs
        vad = self.va_classifier(out["x"])
        logits = self.vap_head(out["x"])

        probs = logits.softmax(dim=-1)

        # 話者間正規化はしない（比較対象の話者がいないため）。
        # 各ビン長で重み付けした発話期待値がそのまま [0, 1] の確率になる。
        p_now = self.objective.probs_self_activity(
            probs,
            from_bin=self.BINS_P_NOW[0],
            to_bin=self.BINS_P_NOW[-1],
        )
        p_future = self.objective.probs_self_activity(
            probs,
            from_bin=self.BINS_PFUTURE[0],
            to_bin=self.BINS_PFUTURE[-1],
        )

        p_bins_tensor = self.objective.probs_bins(probs)

        # p_now / p_future と同じビン範囲の周辺確率を平均して [0, 1] に収める
        i0, i1 = self.BINS_P_NOW[0], self.BINS_P_NOW[-1]
        j0, j1 = self.BINS_PFUTURE[0], self.BINS_PFUTURE[-1]
        p_bins_now_t = p_bins_tensor[:, :, i0 : i1 + 1].mean(dim=-1)
        p_bins_future_t = p_bins_tensor[:, :, j0 : j1 + 1].mean(dim=-1)

        # Get back to the CPU
        p_bins_all = p_bins_tensor.to("cpu").tolist()[0]
        p_bins_now_all = p_bins_now_t.to("cpu").tolist()[0]
        p_bins_future_all = p_bins_future_t.to("cpu").tolist()[0]
        p_now_all = p_now.to("cpu").tolist()[0]
        p_future_all = p_future.to("cpu").tolist()[0]
        vad_all = vad.sigmoid().to("cpu").tolist()[0]

        frames = [
            {
                "p_now": p_now_all[t],
                "p_future": p_future_all[t],
                "vad": vad_all[t][0],
                "p_bins": p_bins_all[t],
                "p_bins_now": p_bins_now_all[t],
                "p_bins_future": p_bins_future_all[t],
            }
            for t in range(x1.shape[1])
        ]
        ret = frames if return_all_frames else frames[-1]

        return ret, new_cache

    def forward_batch(self, x1: Tensor) -> Dict[str, Tensor]:
        """Full-sequence forward without a KV cache (offline batch inference).

        Computes the same values as ``forward(x1, return_all_frames=True)``
        but keeps them as tensors with a batch dimension.

        Returns:
            dict: ``p_now``, ``p_future``, ``vad``, ``p_bins_now`` and
            ``p_bins_future`` of shape (B, T), ``p_bins`` of shape (B, T, n_bins).
        """
        o1 = self.ar_channel(x1)
        out = self.ar(o1["x"])

        probs = self.vap_head(out["x"]).softmax(dim=-1)
        p_bins = self.objective.probs_bins(probs)
        i0, i1 = self.BINS_P_NOW[0], self.BINS_P_NOW[-1]
        j0, j1 = self.BINS_PFUTURE[0], self.BINS_PFUTURE[-1]

        return {
            "p_now": self.objective.probs_self_activity(probs, from_bin=i0, to_bin=i1),
            "p_future": self.objective.probs_self_activity(probs, from_bin=j0, to_bin=j1),
            "vad": self.va_classifier(out["x"]).sigmoid().squeeze(-1),
            "p_bins": p_bins,
            "p_bins_now": p_bins[:, :, i0 : i1 + 1].mean(dim=-1),
            "p_bins_future": p_bins[:, :, j0 : j1 + 1].mean(dim=-1),
        }

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_mono model.

Parameters:

Name Type Description Default
conf Optional[VapConfig]

Configuration object for the model. If None, default VapConfig is used.

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

    Args:
        conf (Optional[VapConfig]): Configuration object for the model.
            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.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,
    )

    # There is no second channel to cross-attend to, so the stereo tower of
    # the two-speaker model is a plain causal GPT here. The layer counts are
    # kept so the depth matches the two-speaker model.
    self.ar = GPT(
        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, num_channels=1
    )

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

encode_audio(audio1)

Encode the raw audio input into a feature representation.

Parameters:

Name Type Description Default
audio1 Tensor

Audio waveform for the input speaker.

 required

Returns:

Name Type Description
Tensor Tensor

Encoded features for the input speaker.
Source code in src/maai/models/vap_mono.py 
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
def encode_audio(self, audio1: torch.Tensor) -> Tensor:
    """Encode the raw audio input into a feature representation.

    Args:
        audio1 (torch.Tensor): Audio waveform for the input speaker.

    Returns:
        Tensor: Encoded features for the input speaker.
    """
    x1 = self.encoder1(audio1)

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

    return x1

forward(x1, cache=None, return_all_frames=False)

Forward pass for the mono VapGPT model.

Parameters:

Name Type Description Default
x1 Tensor

Input audio embedded tensor for the input speaker.

 required
cache dict

Cache of past keys/values.

 None
return_all_frames bool

Return one result dictionary for every input time step instead of only the final step.

 False

Returns:

Type Description
Tuple[dict, dict]

Tuple[dict, dict]: Model outputs and updated cache.
Source code in src/maai/models/vap_mono.py 
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
def forward(
    self,
    x1: Tensor,
    cache: Optional[dict] = None,
    return_all_frames: bool = False,
) -> Tuple[dict, dict]:
    """
    Forward pass for the mono VapGPT model.

    Args:
        x1 (Tensor): Input audio embedded tensor for the input speaker.
        cache (dict, optional): Cache of past keys/values.
        return_all_frames: Return one result dictionary for every input
            time step instead of only the final step.

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

    if cache is None:
        cache = {}

    o1 = self.ar_channel(x1, past_kv=cache.get("ar1"))
    out = self.ar(o1["x"], past_kv=cache.get("ar2"))

    new_cache = {
        "ar1": (o1["past_k"], o1["past_v"]),
        "ar2": (out["past_k"], out["past_v"]),
    }

    # Outputs
    vad = self.va_classifier(out["x"])
    logits = self.vap_head(out["x"])

    probs = logits.softmax(dim=-1)

    # 話者間正規化はしない（比較対象の話者がいないため）。
    # 各ビン長で重み付けした発話期待値がそのまま [0, 1] の確率になる。
    p_now = self.objective.probs_self_activity(
        probs,
        from_bin=self.BINS_P_NOW[0],
        to_bin=self.BINS_P_NOW[-1],
    )
    p_future = self.objective.probs_self_activity(
        probs,
        from_bin=self.BINS_PFUTURE[0],
        to_bin=self.BINS_PFUTURE[-1],
    )

    p_bins_tensor = self.objective.probs_bins(probs)

    # p_now / p_future と同じビン範囲の周辺確率を平均して [0, 1] に収める
    i0, i1 = self.BINS_P_NOW[0], self.BINS_P_NOW[-1]
    j0, j1 = self.BINS_PFUTURE[0], self.BINS_PFUTURE[-1]
    p_bins_now_t = p_bins_tensor[:, :, i0 : i1 + 1].mean(dim=-1)
    p_bins_future_t = p_bins_tensor[:, :, j0 : j1 + 1].mean(dim=-1)

    # Get back to the CPU
    p_bins_all = p_bins_tensor.to("cpu").tolist()[0]
    p_bins_now_all = p_bins_now_t.to("cpu").tolist()[0]
    p_bins_future_all = p_bins_future_t.to("cpu").tolist()[0]
    p_now_all = p_now.to("cpu").tolist()[0]
    p_future_all = p_future.to("cpu").tolist()[0]
    vad_all = vad.sigmoid().to("cpu").tolist()[0]

    frames = [
        {
            "p_now": p_now_all[t],
            "p_future": p_future_all[t],
            "vad": vad_all[t][0],
            "p_bins": p_bins_all[t],
            "p_bins_now": p_bins_now_all[t],
            "p_bins_future": p_bins_future_all[t],
        }
        for t in range(x1.shape[1])
    ]
    ret = frames if return_all_frames else frames[-1]

    return ret, new_cache

forward_batch(x1)

Full-sequence forward without a KV cache (offline batch inference).

Computes the same values as forward(x1, return_all_frames=True) but keeps them as tensors with a batch dimension.

Returns:

Name Type Description
dict Dict[str, Tensor]

p_now, p_future, vad, p_bins_now and
Dict[str, Tensor]

p_bins_future of shape (B, T), p_bins of shape (B, T, n_bins).
Source code in src/maai/models/vap_mono.py 
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
def forward_batch(self, x1: Tensor) -> Dict[str, Tensor]:
    """Full-sequence forward without a KV cache (offline batch inference).

    Computes the same values as ``forward(x1, return_all_frames=True)``
    but keeps them as tensors with a batch dimension.

    Returns:
        dict: ``p_now``, ``p_future``, ``vad``, ``p_bins_now`` and
        ``p_bins_future`` of shape (B, T), ``p_bins`` of shape (B, T, n_bins).
    """
    o1 = self.ar_channel(x1)
    out = self.ar(o1["x"])

    probs = self.vap_head(out["x"]).softmax(dim=-1)
    p_bins = self.objective.probs_bins(probs)
    i0, i1 = self.BINS_P_NOW[0], self.BINS_P_NOW[-1]
    j0, j1 = self.BINS_PFUTURE[0], self.BINS_PFUTURE[-1]

    return {
        "p_now": self.objective.probs_self_activity(probs, from_bin=i0, to_bin=i1),
        "p_future": self.objective.probs_self_activity(probs, from_bin=j0, to_bin=j1),
        "vad": self.va_classifier(out["x"]).sigmoid().squeeze(-1),
        "p_bins": p_bins,
        "p_bins_now": p_bins[:, :, i0 : i1 + 1].mean(dim=-1),
        "p_bins_future": p_bins[:, :, j0 : j1 + 1].mean(dim=-1),
    }

load_encoder(cpc_model)

Load and build the audio encoder for the single input channel.

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_mono.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def load_encoder(self, cpc_model):
    """Load and build the audio encoder for the single input channel.

    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()

    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()