v0.55.1
msitarzewski/agency-agentsv0.55.1Oct 2, 2026by github-actions[bot]
Upgrading?
AI Summary
This release focuses on performance improvements for Text-to-Speech (TTS) by defaulting to Deepgram Aura and fixes a critical bug where bot audio could be dropped during pause-mode barge-in resolution.
Key Highlights
- Deepgram Aura TTS now defaults for `pipecat-bot-py`, reducing time-to-first-audio to 0.52s compared to OpenAI's 0.94s.
- Pause-mode barge-in resolution now correctly re-pushes retained tail audio into the media engine, preventing drops.
- OpenAI TTS configuration now supports a specific `BOT_TTS_MODEL` setting (defaults to `tts-1`).
Full Release Notes
### Changed - **`examples/pipecat-bot-py` now speaks with Deepgram Aura TTS by default** (`BOT_TTS_PROVIDER=deepgram`), reusing the `DEEPGRAM_API_KEY` its speech-to-text already needs, and synthesizing directly at the call's 8/16 kHz. Measured on 12 requests each: time to first audio was a median of 0.52 s (worst 0.59 s), against 0.94 s with a 12.7 s worst case for OpenAI `gpt-4o-mini-tts`. On PSTN calls the median gap between the caller's last word and bot audio went from 1.95 s to 1.01 s, and a goodbye that had waited 23 s started in 0.13 s. `BOT_TTS_PROVIDER=openai` remains available, now with a `BOT_TTS_MODEL` setting that defaults to `tts-1` (median 1.64 s, worst 2.3 s) rather than the occasionally stalling `gpt-4o-mini-tts`. ### Fixed - **Pause-mode barge-in no longer drops bot audio after a resolution** ([#620](https://github.com/thevoiceguy/siphon-ai/issues/620)). Resolving a pause arbitration (reject, confirm, or either timeout) re-pushes the retained tail and the post-pause audio straight into the media engine. A WS server still streaming at real time, which is what a correctly paced server does (including `sdks/python`'s `AudioSender` and `sdks/typescript`), then accumulated behind that backlog, and the 200 ms window (PROTOCOL.md §5.5) evicted about the pause length minus 200 ms of bot speech per resolution: 158 frames over 5 rejects on a live PSTN call. The window now tolerates the backlog the daemon itself created, shrinking as it drains and resetting on every flush, so PROTOCOL.md §4.11's "playout resumes where it stopped, then any audio streamed during the pause plays behind it" holds. Over-rate servers are still bounded. No configuration or protocol change. --- **Artifacts** - Static musl binaries for `x86_64` and `aarch64` (`.tar.gz`). - Debian packages (`.deb`) for `amd64` and `arm64` — `apt install ./siphon-ai_*_amd64.deb`. - `SHA256SUMS` — verify with `sha256sum -c SHA256SUMS`. - `*-sbom.cdx.json` — CycloneDX SBOM. - `SHA256SUMS.cosign.bundle` — cosign keyless signature over the checksums. - Container: `ghcr.io/thevoiceguy/siphon-ai:v0.55.1` (multi-arch, cosign-signed). Verify the checksums signature: ``` cosign verify-blob --bundle SHA256SUMS.cosign.bundle \ --certificate-oidc-issuer https://token.actions.githubusercontent.com \ --certificate-identity-regexp 'release.yml@' SHA256SUMS ```
Related Repositories
obra/superpowers★ 303,157
An agentic skills framework & software development methodology that works.
0xNyk/council-of-high-intelligence★ 4,687
Structured multi-perspective deliberation for hard decisions. Run full councils, focused triads, or duo debates across Claude Code, Codex, Gemini CLI, and OpenCode.
LukeSmithxyz/emailwiz★ 2,171
Script that installs/configures a Dovecot, Postfix, Spam Assassin, OpenDKIM Debian web server
dmachat/angular-webpack-cookbook★ 346
The repository for the Angular Webpack Cookbook Gitbook
shivangi-aneja/FaceTalk★ 239
[CVPR 2024] FaceTalk: Audio-Driven Motion Diffusion for Neural Parametric Head Models