Voice AI's money story arrives: record revenue and a real-time model
SoundHound reported Q3 revenue of $62.4 million, up 48.5% year over year, and guides to sustained adjusted EBITDA profitability within the second half of 2027. Its LivePerson acquisition closed on September 4, 2026.
- SoundHound AI reported third quarter 2026 revenue of $62.4 million, a 48.5% year-over-year increase, and a GAAP net loss of $46.4 million. m:…6936
- SoundHound completed the acquisition of LivePerson on September 4, 2026, and identified $120 million in annualised cost synergies, with $80 million expected to take effect before January 1, 2027. m:…6936
- SoundHound expects to reach sustained adjusted EBITDA profitability within the second half of 2027, and projects 2027 revenue of $350 million to $400 million. m:…6936
- Boson AI's Higgs Realtime Duplex is a 45B mixture-of-experts model that processes audio and video in 160 ms chunks and scored 98.9 on Big Bench Audio accuracy. Self-reported m:…1412
- Superwhisper is now the default voice input on Omarchy, pre-installed on new x86_64 installations. Self-reported m:…0765
The quarter
SoundHound AI reported third quarter 2026 revenue of $62.4 million, a record for the company and a 48.5% increase year over year. The company reported a GAAP net loss of $46.4 million for the quarter, and a GAAP gross margin of 34.1%. It had $207.6 million in total cash and cash equivalents as of September 30, 2026. The company expects 2026 revenue in the range of $250 million to $260 million, and projects 2027 revenue of $350 million to $400 million. m:…6936
The more important line is the one about profitability. SoundHound says it expects to reach sustained adjusted EBITDA profitability within the second half of 2027, and it is shifting from upfront licence deals to hosted structures to improve earnings quality. It is also strategically exiting non-core, dilutive business lines and exiting negative gross margin contracts inherited from prior acquisitions. The company is cleaning up the revenue mix while it grows it, which is what a path to profitability actually looks like. m:…6936 m:…3422
The acquisition
SoundHound closed its acquisition of LivePerson on September 4, 2026, and the integration math is now public. The company has identified $120 million in annualised cost synergies, with $80 million expected to take effect before January 1, 2027. That is a large number relative to a $62.4 million quarter, and it explains the shape of the deal: buy a customer-service business with scale, cut the overlap, and keep the voice AI platform on top. m:…6936
The product news underneath the deal is about where voice goes next. SoundHound launched OASYS Edge to bring LLM-powered voice AI agents to local device hardware, and signed a deal with LG to debut a commerce channel across approximately 20 million smart TVs in the US. Voice on a TV is a different product from voice in a car or a call centre: the screen is the storefront, and the voice is the salesperson. Whether shoppers talk to their television is still an open question, but SoundHound is now betting on it at scale. m:…3422
The model that talks back
Boson AI introduced Higgs Realtime Duplex, a full-duplex interactive omni model that listens, watches and talks at the same time while performing background tasks. The front model is a 45B mixture-of-experts model, trained on approximately 2 trillion tokens of text, audio and video, and it processes audio and video in 160 ms chunks to maintain natural conversational timing. It scored 98.9 on Big Bench Audio accuracy, according to the company, and autonomously handles reasoning and tool calls without delegating to a background model. m:…1412
The economics are the interesting part. Boson AI says it aims for an hour of model usage to cost at most one-tenth of an hour of human time, and the model uses roughly 2.6K to 5K tokens per minute depending on workload. Voices and faces are loaded zero-shot at deployment time using reference clips and images, and the model retains the last five minutes of audio and visual streams in production. An API is planned for a future release; Higgs Realtime is available today as an end-to-end model with avatar output. The company trained the model on its own NVIDIA H100 cluster in a data centre it built itself. m:…1412
Voice at the operating-system level
Superwhisper's announcement is smaller but points in the same direction. The company partnered with DHH to become the default voice solution on Omarchy, the Linux desktop distribution, and Superwhisper comes pre-installed on new x86_64 installations. Omarchy manages recording shortcuts and loads the integration automatically, without manual Hyprland configuration. The first 10,000 words transcribed via Superwhisper in the cloud are free, and users can switch to a free downloadable model that runs entirely on their machine. m:…0765
Put together, the three stories describe voice AI moving down the stack: a public company reporting real revenue and a path to profit, a model that talks and listens in real time, and a voice layer shipping as a default in an operating system. The demo era for voice is over. The question now is whether the unit economics — SoundHound's hosted margins, Boson's tenth-of-a-human-hour target, Superwhisper's free local model — hold up when the usage is real and continuous rather than a conference demo. m:…6936 m:…1412 m:…0765
What is still unresolved
- Whether SoundHound's $120 million in synergies materialises on schedule.
- Whether shoppers actually use a voice commerce channel on a smart TV.
- When Boson AI ships the Higgs Realtime Duplex API, and at what price.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.