Speaker embedding¶

On-device speaker embedding for enrollment and cosine verification. Public type: SpeakerEmbedding (family thestage_speaker_embedding).

Produces a 192-d L2-normalized vector from 16 kHz mono PCM. Default compute device is CPU.

Basic usage¶

Swift (singleton JSON path — same shape as Flutter):

try await TheStageAI.shared.start_model(
    model_name: "speaker_id",
    engines_path: "TheStageAI/redimnet2",
    model_type: "thestage_speaker_embedding"
)

let out = try TheStageAI.shared.infer(
    model_name: "speaker_id",
    input_json: ["audio": pcm_16k_mono]
)
// out[0]["embedding"] → [Double] length 192

let check = try TheStageAI.shared.infer(
    model_name: "speaker_id",
    input_json: [
        "audio": pcm_16k_mono,
        "embedding": reference_embedding,
    ]
)
// check[0]["similarity"] → Double (cosine)

Flutter:

await TheStageFlutterSDK.start_model(
  model_name: 'speaker_id',
  engines_path: 'TheStageAI/redimnet2',
);

final enrolled = await TheStageFlutterSDK.infer(
  model_name: 'speaker_id',
  input_json: {'audio': pcm16k},
);

Audio contract¶

Item	Value
Sample rate	16 000 Hz mono Float / `Float32List`
Window	2.0 s (32 000 samples) — pad/trim trailing
Output dim	192, L2-normalized
Device default	`cpu`

Caller chooses the accept threshold (voice agent default is 0.75).

Voice agent¶

Set speaker_id on the agent config and call enroll_speaker. See Voice Agent.

Agent checklist¶

Family id: thestage_speaker_embedding; common handle: speaker_id.
Always 16 kHz mono; 2 s window.
Verify = audio + reference embedding → similarity.

Speaker embedding