Familiar

{"summary_one_liner":"Familiar builds a full-duplex speech AI model for revenue and sales calls that listens while it speaks.","detailed_description":"Familiar develops a duplex speech model designed for revenue and sales calls. Rather than assembling a cascaded pipeline of automatic speech recognition, text-based reasoning and text-to-speech, the company offers a single model that it says listens, thinks and speaks within 350 milliseconds and keeps both audio channels open, so that interruptions, overlapping speech and background voices do not break the conversation. The model reasons over speech rather than transcribed text, taking into account not only the words spoken but how they were said and what is happening in the surrounding audio environment, and is described as asking clarifying questions when uncertain.\n\nThe product is positioned against what the company characterises as the failure modes of existing voice agents: \"walkie-talkie\" turn-taking, awkward pauses, cut-offs, and transcription-only understanding that loses detail. Familiar's site cites a claim that over 40% of people hang up on voice agents within 30 seconds, framing each hang-up as a lost customer. Customers do not need to build their own ASR, TTS, turn detector or dialogue harness; instead they define the workflow via a prompt, supply tools for the model to call, and select a personality.\n\nOperationally, Familiar offers guardrails that can filter or block a response before the caller hears it, per-turn debugging across reasoning, tool calls and speech (in both text and audio), and streaming of telemetry into a customer's existing evaluation and monitoring tooling. The model can be improved over time through feedback coaching or fine-tuning on a customer's own calls."},"technology":"Familiar's core technology is a single full-duplex speech-to-speech model that keeps both the listening and speaking channels open simultaneously, allowing a caller to barge in mid-sentence while the model yields, continues listening and resumes. The company states an end-to-end listen-think-speak latency of 350 ms and says the model learned the mechanics of conversation directly from human data. It reasons over the speech signal itself rather than a text transcript, capturing prosody and ambient context in addition to words, and replaces the conventional cascaded stack of ASR, TTS, turn detection and dialogue orchestration with one model. Production features include response guardrails, layer-by-layer inspection of each turn's reasoning, tool calls and speech in both text and audio, and streaming into existing evaluation and monitoring systems.","business_model":"Familiar sells its speech model to businesses that operate voice agents, offering a dedicated deployment rather than a shared API. The company positions cost as being at parity with a customer's existing cascaded voice stack. The stated entry point is a 30-day pilot on a single use case.","target_customers":"Organisations running voice agents on revenue and sales calls, including those with data-residency or privacy requirements that need on-premises or private-cloud deployment.","go_to_market":"The company's website markets a 30-day pilot on one use case as the initial engagement, and includes an interactive click-to-speak demo plus an unedited recording of a real call demonstrating how the model handles interruptions and overlapping speech.","differentiation":"Familiar differentiates on full-duplex conversation handling versus half-duplex, turn-based voice agents built on cascaded pipelines; on reasoning directly over speech instead of text transcripts; on a single model that removes the need to integrate ASR, TTS, turn detection and a dialogue harness; and on deployment inside the customer's own VPC or on-premises so that data does not leave the customer's network, with claimed cost parity to a cascaded stack.","geographic_footprint":"Headquartered in London, United Kingdom (internal record); the public sources reviewed do not specify additional locations.","latest_developments":"The company's website describes a shipping product with an interactive demo, on-premises availability and an open 30-day pilot offer; no dated announcements were available in the sources reviewed."}