语音与电话

Real-time voice input and output. Four channels cover the matrix: inbound calls, local microphone wake, outbound speech synthesis, and SIP-grade real-time conversation.

ClawdTalk（实时 SIP）

[channels.clawdtalk]
enabled = true
api_key = "..."                                # Telnyx API key (secret)
connection_id = "..."                          # Telnyx SIP connection ID
from_number = "+14155550123"                   # caller-ID for outbound dials
allowed_destinations = ["+14155551234"]        # destinations allowed for outbound dial; empty = none
webhook_secret = "..."                         # optional: shared secret for inbound Telnyx webhook verification

Full-duplex SIP voice powered by Telnyx. The agent talks over a real phone call (inbound or outbound). Supports barge-in, mid-turn tool use, and regional number provisioning.

Pair with: a telnyx model provider for the brain (crates/zeroclaw-providers/src/telnyx.rs) and ensure your Telnyx account has a SIP connection with the correct webhook URL pointed at the ZeroClaw gateway.

语音通话（Twilio / Telnyx / Plivo）

[channels.voice_call]
enabled = true
provider = "twilio"                            # "twilio" (default), "telnyx", or "plivo"
account_id = "..."                             # provider-specific account identifier
auth_token = "..."                             # provider-specific auth token (secret)
from_number = "+14155550123"
webhook_port = 8090                            # default 8090; embedded webhook server
require_outbound_approval = true               # default true; require operator approval before dialing
transcription_logging = true                   # default true; persist call transcripts
# tts_voice = ""                               # optional voice ID override (provider-specific); omit to use provider default
max_call_duration_secs = 3600                  # default 3600 (1 hour cap)
# webhook_base_url = ""                        # optional public base URL when behind a tunnel/proxy; omit to use the localhost fallback

Traditional carrier voice — the agent picks up, transcribes the caller, replies with TTS. Higher latency than ClawdTalk but works with any regular phone number and doesn’t require SIP trunk provisioning. Outbound calls hit from_number and require operator approval when require_outbound_approval is on.

语音唤醒（本地唤醒词）

[channels.voice_wake]
wake_word = "hey zeroclaw"                     # default "hey zeroclaw" (case-insensitive substring match)
silence_timeout_ms = 2000                      # default 2000; ms of silence before finalising capture
energy_threshold = 0.01                        # default 0.01; RMS energy below this is treated as silence
max_capture_secs = 30                          # default 30; hard cap on capture duration

在本地运行，监听麦克风，在听到唤醒词时触发代理交互。适用于：

• SBC 上的物理语音助手
• 桌面端“热词 → 提问”工作流
• 始终在线的家庭自动化代理

The agent doesn’t send audio anywhere — wake detection is local. Only post-wake speech is captured and (separately) transcribed before reaching the LLM.

Build flag: Voice Wake is gated by the voice-wake cargo feature on zeroclaw-channels. Build with --features voice-wake to include it.

TTS（出站语音合成）

TTS lives at the top level under [tts], not under [channels.*] — it’s an output service that channels can call into, rather than its own inbound channel.

[tts]
enabled = true
default_provider = "piper"                     # "openai", "elevenlabs", "google", "edge", or "piper"
default_voice = "en_US-lessac-medium"          # provider-specific default voice ID
default_format = "mp3"                         # "mp3" (default), "opus", or "wav"
max_text_length = 4096                         # default 4096

[tts.openai]
api_key = "..."
model = "tts-1"                                # default "tts-1"
speed = 1.0                                    # default 1.0

[tts.elevenlabs]
api_key = "..."
model_id = "eleven_monolingual_v1"             # default "eleven_monolingual_v1"
stability = 0.5                                # default 0.5
similarity_boost = 0.5                         # default 0.5

[tts.google]
api_key = "..."
language_code = "en-US"                        # default "en-US"

[tts.edge]
binary_path = "edge-tts"                       # path to the edge-tts binary; default "edge-tts"

[tts.piper]
api_url = "http://127.0.0.1:5000/v1/audio/speech"   # OpenAI-compatible Piper HTTP endpoint

Only the section for the active default_provider needs to be filled in. Pair [tts] with voice_wake for a complete local voice assistant.

延迟预算

端到端延迟低于约 500 毫秒时，语音听起来是实时的。实际预算：

ClawdTalk 通过保持音频流实时传输，简化了其中几个步骤；而常规的 voice_call 则依次经历 STT + LLM + TTS。

STT

Speech-to-text is configured separately from the voice channels — see the [transcription] config in the Config reference. Voice channels invoke whichever transcription provider is active when they need to turn audio into text.

硬件说明

对于始终开启的语音功能，在 SBC 上：

• USB mic: any UAC-compliant mic works. arecord -l to verify the OS sees it.
• Speaker: either USB audio out or the SBC’s onboard jack; pick the OS default device for the user the daemon runs as.
• Microphones with built-in AEC (acoustic echo cancellation) dramatically improve wake reliability when the speaker is nearby.

有关 Android 特定的音频设置，请参阅 硬件 → Android。