Voice AI platforms face scrutiny over performance claims at scale, Deepgram says
The speech AI company said independent benchmarks testing platforms at thousands of concurrent connections do not exist
Deepgram, the artificial intelligence speech platform, has said voice AI systems often degrade significantly when handling thousands of concurrent calls, with no independent production-scale benchmarks available to verify vendor claims.
The company cited research from Teneo.ai, the conversational AI firm, estimating that speech recognition failures cost US contact centres $934 million annually.
Deepgram said no publicly available tests evaluate platforms at 5,000 or more concurrent connections, leaving engineering teams to rely on private proof-of-concept testing to validate performance.
The company highlighted wide variation in pricing transparency across the market, with per-minute rates ranging from $0.0025 to $0.024 and credit-based models translating to roughly $0.05 per minute, creating cost differences of two to twenty times at scale.
Deepgram's evaluation assessed several providers, positioning itself as a leading business-to-business infrastructure and citing customer results, including a reported two to four times improvement in alphanumeric transcription accuracy for Five9 and up to 89.6% accuracy on NASA space-to-ground communications.
AssemblyAI was noted for the lowest documented per-minute pricing at $0.0025 and SOC 2 Type II certification, while Google Cloud Speech and Amazon Web Services Transcribe were identified as the only providers holding FedRAMP High authorisation, the US government's top-tier cloud security standard.
The evaluation found that Microsoft Azure offers batch pricing advantages but has documented SDK stability issues, while integration complexity across providers ranges from hours for WebSocket software development kits to weeks for those requiring complex authentication.
Related reading
- Saudi mega-project Qiddiya turns to Microsoft AI to manage construction complexity
- End-to-end speech models cut voice latency by up to 70%, Deepgram says
- Google Cloud builds AI video tool for US Ski & Snowboard athletes ahead of Olympics
Among other vendors assessed, ElevenLabs reported approximately 75 millisecond inference latency, WellSaid Labs published a 99.99% uptime service-level agreement, and Murf documented support for 10,000 simultaneous calls.
Deepgram recommended that engineering teams conduct private proof-of-concept tests using production-representative audio at target concurrency levels, and said it offers $200 in free credits for testing.
The Recap
- Independent production-scale voice AI benchmarks do not exist publicly.
- Per-minute pricing ranges from $0.0025 to $0.024 per minute.
- Google Cloud Speech and AWS Transcribe hold FedRAMP High.
