SoundHound's Q2 Revenue Grew 45%. Its Voice AI Output Quality Is Still Unmeasured.

SoundHound AI posted record Q2 2026 revenue of $61.9 million, a 45% increase year-over-year. The company raised the low end of its full-year guidance to $230 million. Shares rose 15% in the days following the August 5 announcement. CEO Keyvan Mohajer teased upcoming developments in in-house LLMs and text-to-speech technology.

The company also recorded a net loss of $42.8 million. After nearly two decades of operation, SoundHound has never turned a profit. And its stock has lost roughly 25% of its value over the past 12 months, even as revenue accelerated.

Revenue is up. The market is not convinced. That gap deserves attention.

What does SoundHound's Q2 2026 reveal about voice AI at scale?

SoundHound's earnings report is a public window into a pattern that runs through the entire voice AI industry: platform adoption metrics grow while output quality metrics remain absent.

The Q2 earnings call covered revenue growth, customer counts, deployment expansion across automotive, healthcare, and restaurants, and the company's new voice-native agentic AI platform that pulls in models from other providers. Every metric reported describes how the platform performs. None describe whether the audio reaching end-users is correct.

Revenue counts calls processed. It does not count mispronounced customer names, drifted voice profiles, or audio that failed telephony format requirements. A call that processes successfully and a call that mispronounces the caller's account number both contribute the same revenue.

Meanwhile, ElevenLabs reported a $600 million annual revenue run rate in July 2026, already more than double SoundHound's projected full-year revenue. The generation side of voice AI is growing faster than ever. The validation side has not kept pace.

Why does 45% revenue growth not translate to profitability?

SoundHound has now operated for nearly 20 years without achieving profitability. The company holds $203 million in cash and no debt, but the $42.8 million quarterly loss signals that scaling deployments does not automatically scale margins.

Part of the answer sits in the cost structure that voice AI platforms rarely disclose. When voice outputs contain pronunciation errors, inconsistent voice profiles, or format failures, the costs show up downstream: customer complaints, manual re-recordings, engineering hours spent debugging quality issues that no automated system caught, and client churn that shortens contract lifetime value.

Forrester's Q2 2026 Wave report found that voice AI handles 19% of inbound contact center volume in 2026, up from 6% in 2024. That growth rate means the volume of unvalidated voice output is roughly tripling every two years. At SoundHound's scale, with 25 languages and specialized enterprise deployments, each language and each vertical is a separate quality surface. Every new deployment adds pronunciation dictionaries, voice profiles, and format requirements that the platform must get right on every output.

A 2% pronunciation error rate sounds small. At enterprise call volume, 2% means thousands of wrong outputs per quarter. Those outputs carry real costs that revenue metrics never capture.

What metrics are missing from voice AI earnings reports?

Every publicly traded voice AI company reports the same category of metrics: revenue, customer count, call volume, containment rate, average handle time, CSAT scores. These are agent-level and platform-level measurements. They describe the system's decisions and its commercial traction.

No voice AI earnings report includes:

• Per-output pronunciation accuracy on domain vocabulary (drug names, account numbers, street addresses, product SKUs)
• Model version tracking across deployments, confirming every production instance runs the validated version
• Voice consistency scoring, measuring whether the same voice profile sounds identical across a library of thousands of outputs
• Audio format compliance rates for telephony infrastructure (G.711 codec, 8kHz sample rate, EBU R128 loudness normalization)

These are output-level measurements. They describe whether the audio the end-user hears is correct. The fact that no public voice AI company reports them tells you they either do not measure them or the numbers are not flattering.

According to a McKinsey analysis, AI resolutions cost $0.62 per interaction versus $7.40 for human agents. That 12x cost advantage only holds if the AI resolution is actually correct. A resolution that mispronounces a loan balance or reads back the wrong callback number is a resolution on paper and a failure in practice.

How does a production layer close the gap between revenue and quality?

The gap between revenue growth and profitability in voice AI is partly a validation gap. Platforms that generate audio without systematically validating every output before delivery carry hidden quality costs that compound with scale.

Onepin is a voice workflow platform that orchestrates, validates, and ships production-ready audio across 100+ TTS models. Instead of measuring success by whether audio was generated, it measures success by whether audio is correct.

The production pipeline works in four stages:

1. Route each request to the best model for the language, voice profile, and use case
2. Generate the audio using the pinned model version
3. Validate every output against the locked reference profile, scoring pronunciation accuracy, voice consistency, and format compliance
4. Regenerate only the clips that fail, using targeted retry logic instead of re-running the entire batch

This approach converts quality from an assumption into a measurement. Every output ships with a quality score, a model version stamp, and an audit trail. When a deployment scales from 1,000 to 100,000 outputs, the validation layer scales with it.

Revenue growth measures adoption. Output quality measures whether adoption creates value or creates liability. SoundHound's Q2 shows you can have the first without proving the second.

Build the validation layer at onepin.ai.