Repository navigation
Releases
Releases · amoh-ai/harbor-voice
v0.1.2
harbor-voice, harbor-voice-tau-voice-adapter and harbor-voice-eva-adapter 0.1.2, on PyPI:
pip install "harbor-voice[tau-voice,eva]==0.1.2"New
- More providers. Speech-to-speech: Google Gemini Live (
google.live), xAI Grok (xai.realtime), Deepgram Voice Agent (deepgram.agent), and a self-hosted realtime server (selfhosted.realtime). STT and TTS: Deepgram, Cartesia, Fish Audio, Sarvam, Microsoft Azure STT, Meta STT, xAI and OpenRouter. LLMs: LiteLLM, OpenRouter and self-hosted. - Cascade agents on tau-voice. An STT, an LLM and a TTS can be the agent under test (
tau-voice job-config --cascade), as well as a speech-to-speech model. - OpenAI Responses API.
api = "responses"onopenai.llm, with reasoning carried across tool rounds. - x86 and ARM images, datasets 0.1.1. Every image is built for
linux/amd64andlinux/arm64. Datasets 0.1.1 pins them. - Omni + TTS cascade. An omni model hears the caller and writes the reply, and a TTS speaks it:
selfhosted.omni(vLLM) andopenrouter.omni. Runs on the built-in call loop. - Bedrock chat.
aws.bedrock-llmholds a whole conversation (Converse), so it can drive the tau-voice cascade. - OpenAI instructions.
systemandsystem_role(system,developer, or the Responses API'sinstructions) onopenai.llm. - Self-hosted providers. Speech and LLM servers behind your own URL, Modal included.
- Persona caller. It gets its brief, hears the agent (through an optional
[caller] sttfor a speech-to-speech agent), and can end the call with[END CALL]. - EVA judge routes.
eva-adapter judges --gemini api-key|vertex|vertex-json: Gemini on the API, on Vertex with ADC, or on Vertex with a service account (VERTEXAI_CREDENTIALS).
Fixed
- Full duplex waits 1 s of quiet before ending a turn, so replies are no longer cut off. The silence after a reply is no longer counted as part of it, or as bleed.
- A failing cascade slot ends the call at once, instead of waiting out every turn's timeout or going unreported after the last turn.
- Provider keys in vendor headers can't follow a redirect to another host. Redirects are followed only by self-hosted servers.
- Bedrock chat no longer sends a blank system text. A non-JSON 200 reply now raises a clear MALFORMED error.
- The persona caller's lines stay paired with the right agent replies after probes. A second run starts clean. Clients are closed when a call ends.
- The keys file in the sandbox is removed even if loading it fails.
- Deepgram TTS retries a dropped stream before any audio was sent.
Datasets 0.1.2
All eight datasets are registered at harbor-voice-datasets v0.1.2. Only the EVA tasks change: they pass VERTEXAI_CREDENTIALS to the sidecar. The tau-voice tasks are the same as 0.1.1.
eva-adapter job-config --hub-dataset eva@0.1.2 --model openai.realtime/gpt-realtime-1.5 -o eva.yamlThe images are unchanged from 0.1.1, built for amd64 and arm64:
| Image | Tag |
|---|---|
| tau-voice sidecar | ghcr.io/amoh-ai/tau-voice-runtime:2174a60-ce6c029644 |
tau-voice main |
ghcr.io/amoh-ai/tau-voice-main:413a7f6ab7 |
| EVA sidecar | ghcr.io/amoh-ai/eva-sidecar:039efa0a-077f8e4f23 |
EVA main |
ghcr.io/amoh-ai/eva-main:413a7f6ab7 |
Housekeeping
- The package ships
py.typed. CI also tests Python 3.12. - The lock file is refreshed: filelock 4.2.10, dnspython 2.9.2, opentelemetry 1.45.2. This clears the open Dependabot alerts.
- Added
SECURITY.md(report privately through the Security tab) and a weekly Dependabot config.
Pull requests
- update: eva parity, by @uditmpradhan in #3
- More providers, built-in call loop fixes, release 0.1.2, by @darkoether in #4
- Test: stop the dither check flaking on short cases, by @darkoether in #5 (after the tag)
Full changelog: v0.1.1...v0.1.2
v0.1.1
What's Changed
- Fix L2 speaking rate and dynamic muffling; release 0.1.1 by @uditmpradhan in #2
Full Changelog: v0.1.0...v0.1.1
v0.1.0
What's Changed
- PyPI release: install from PyPI, release checks and workflow by @uditmpradhan in #1
New Contributors
- @uditmpradhan made their first contribution in #1
Full Changelog: https://github.com/amoh-ai/harbor-voice/commits/v0.1.0