SpatialReal Python SDK

pip install spatialreal — realtime avatar sessions against the SpatialReal backend: stream audio in, get lifecycle signals back, and (in egress mode) have the avatar's synchronized audio/video published straight into a LiveKit room or Agora channel.

Usage

from datetime import datetime, timedelta, timezone
from spatialreal import LiveKitEgressConfig, new_avatar_session

session = new_avatar_session(
    api_key="...",
    app_id="...",
    avatar_id="...",
    console_endpoint_url="https://api.spatialreal.cloud",  # OpenAPI root: session tokens
    ingress_endpoint_url="wss://driven.us-west.spatialreal.cloud/v2/driveningress",
    expire_at=datetime.now(timezone.utc) + timedelta(hours=1),
    sample_rate=16000,
    livekit_egress=LiveKitEgressConfig(url=..., api_token=..., room_name=..., publisher_id=...),
    on_playback=lambda sig: print(sig.req_id, "ended" if sig.end else "playing"),
    on_error=lambda err: print("error:", err, "retryable:", getattr(err, "retryable", None)),
    on_close=lambda: print("closed"),
)
await session.init()  # API key -> session token (console API)
await session.start()  # WebSocket + handshake; read loop starts

req_id = await session.send_audio(pcm_bytes)  # same id until end=True
await session.send_audio(b"", end=True)  # closes the segment
await session.interrupt()  # interrupts the latest request
await session.close()

Pause / resume (egress mode)

When the server declares the playback_control capability, the current segment can be paused and resumed instead of interrupted — the server holds its send cursor with zero regeneration:

await session.pause()  # server stops writing frames, keeps ingesting audio
# ...keep sending audio, including end=True, while paused...
await session.resume()  # continues from where it stopped

Feature-detect with "playback_control" in session.capabilities. Subscribe to on_playback_state(PlaybackStateEvent) for the resulting state (PLAYING / PAUSED / ENDED / INTERRUPTED with played_ms and, on interrupt, a reason such as pause_timeout). Older servers never send it and ignore pause/resume.

Session objects are single-use: on a dropped connection, create a fresh one (request state is deliberately not reusable across connections). AvatarSDKError.retryable tells you whether reconnecting can succeed — it implements the server's WebSocket close-code contract (40xx: don't retry, 45xx: retry with backoff).

What's new

• on_playback(PlaybackSignal) — structured playback lifecycle; no protobuf parsing in caller code (transport_frames(raw, is_last) still exists).
• CloseCode + AvatarSDKError.retryable — the reconnect contract as API.
• session.capabilities — server-declared capabilities from the handshake (feature-detect, don't version-detect).
• pause() / resume() + on_playback_state(PlaybackStateEvent) — egress-mode server-side playback control (gated on the playback_control capability).
• on_close fires exactly once per session; the token request has a timeout.
• AudioFormat.OGG_OPUS takes pre-encoded bytes; the SDK does not encode client-side.

Development

uv venv && uv pip install -e ".[dev]"
python -m pytest tests -q     # protocol-level tests against an in-process fake backend
./scripts/gen_proto.sh        # regenerate protobuf from proto/ (see proto/SHARED_PROTO_COMMIT)

License

Apache-2.0