AI 推論與語音代理
Pipecat 1.10 Overhauls Voice Agent Provider Interfaces; OpenAI 3 Upgrade May Change TLS Behavior
Pipecat 1.10 adds cross-chunk speech continuity, server-side utterance segmentation, and configurable audio buffering, while supporting newer OpenAI, Anthropic, and MCP SDKs. The release also changes some HTTP, STT, and inference defaults, meaning compatibility of existing containers and voice pipelines cannot be determined through package-level testing alone.
The open-source real-time voice agent framework Pipecat released version 1.10.0 on September 12. Smallest TTS now uses the same `context_id` to link text chunks within a single LLM turn, preventing prosody from resetting with every streaming request. Server-side buffering can also be controlled using `max_buffer_delay_ms`. Gradium STT adds server-side turn detection and endpointing parameters such as `eot_horizon_s` and `eot_threshold`. The LiveKit output queue can now be configured through `audio_out_queue_size_ms` to reduce underruns on high-jitter networks.
Of greater concern to deployment teams are dependency and default-behavior changes. The supported `openai` version range now extends to 3.x. When OpenAI 3 is resolved, the underlying client switches to `httpx2`, while TLS certificate verification moves from `certifi` to the operating system trust store. Minimal containers without system CA certificates, or enterprise environments using TLS inspection, may need to configure `SSL_CERT_FILE` or `SSL_CERT_DIR`. The `Timeout` type passed to services must also match the HTTP client family actually in use. The Anthropic SDK can be upgraded to 1.x, and the MCP client can work with either the 1.x or 2.x SDK, but custom Bedrock clients must now explicitly provide an AWS region.
The Speechmatics changes are closer to a migration: the service now connects to `/v2/agent`, uses the provider's VAD for turn detection by default, and removes speaker focus along with several legacy parameters. Meanwhile, the DeepSeek service disables thinking by default to reduce time to first token in voice interactions and fixes a 400 error caused by missing `reasoning_content` after tool calls. Engineering teams should use real calls to remeasure time to first word, barge-in behavior, reconnection, and the ordering of end events. The release notes provide no end-to-end latency or stability benchmarks, so the new buffering controls should not be treated as a performance improvement by default.