[HN Gopher] Show HN: Vogent - Better Building Blocks for Voice AI
___________________________________________________________________
Show HN: Vogent - Better Building Blocks for Voice AI
Hi HN! Excited to share some stuff we've been building. We spent
the last year building voice agents to automate individual call
tasks for companies with large call centers. The STT-LLM-TTS-VAD
cycle is mostly solved at this point, but the last-mile problem for
making these agents performant was frustrating. We ended up
building a lot of band-aids, and we decided to put them together
into their own end-to-end product. Vogent is a platform for
building and serving Voice AI agents, with a focus on providing
higher-level building blocks that make it easy to get a voice agent
working quickly. You can check out the docs at
https://docs.vogent.ai It supports the typical design process of a
voice agent (choosing/prompting a model, selecting a voice, and
hosting on a phone number or accessing via API), but it has
additional pieces that make voice agents performant quickly, like
(among other things): - A drag-and-drop agent builder Under the
hood, this involves feeding the model only context relevant to the
goals of the current node (e.g., asking a particular question and
probing conversationally for the answer), while giving it the
ability to call a function with the outcome once the goal is
achieved to transition to the appropriate next node. This makes it
easy to build voice agents that need the structure of a multi-step
talk track with the flexibility of accomplishing each task
conversationally. - Voices that are trained to spell Off-the-
shelf voices (e.g. Eleven, Cartesia) sound _much_ more artificial
when they spell. It might sound like an edge case, but this killed
almost every engagement we had. We ended up recruiting Upworkers
with different accents, having them spell a few thousand phrases,
and training our own voices by modifying open-source architectures.
Choose "Carlos" for a spelling-optimized voice right now; we're
adding a lot more soon. - An IVR detection model This detector
uses the audio stream to predict whether a line came from an IVR or
a human, and switches between different LLMs based on the result
(so you can have independent IVR navigation and conversational
models). - Model versioning and counterfactuals Vogent enables
model versioning and testing against past dials within the product.
Any feedback would be appreciated. Please also feel free to join
our Discord: https://discord.gg/JmThYcyG
Author : jag729
Score : 15 points
Date : 2025-01-29 16:39 UTC (6 hours ago)
(HTM) web link (www.vogent.ai)
(TXT) w3m dump (www.vogent.ai)
| CharlesW wrote:
| > _Please also feel free to join our
| Discord:https://discord.gg/JmThYcyG_
|
| FYI, I get an "Invalid Invite" error.
| jag729 wrote:
| Oh weird, let me take a look at that
| dvaun wrote:
| I keep seeing similar aesthetics for landing pages and startup
| sites. Is there some common template that's become popular for
| this purpose?
___________________________________________________________________
(page generated 2025-01-29 23:01 UTC)