June 2025 Summaries

Choosing among Deepgram, OpenAI, and Google speech-to-text APIs requires balancing transcription accuracy, latency, cost, formatting quality, specialized features, and deployment needs rather than selecting a universally best provider. Word Error Rate is a common accuracy metric, but results vary depending on whether text is normalized, whether audio is processed in batch or streaming mode, and the dataset’s length, noise, accents, languages, and technical vocabulary. Third-party and provider benchmarks generally suggest that OpenAI’s GPT-4o Transcribe often performs strongly on raw and formatted batch-transcription accuracy, Deepgram Nova-3 offers particularly high streaming speed, competitive pricing, and strengths with technical or challenging audio, while Google Chirp 2 can deliver solid accuracy but has a higher standard price. Streaming typically sacrifices some accuracy and formatting quality for real-time responsiveness, whereas batch processing benefits from more context. The comparison emphasizes that vendor benchmarks can use differing methodologies and may not reflect an application’s actual audio, so developers should test representative samples, assess both raw and presentation-quality errors, account for feature requirements such as diarization and custom vocabulary, and calculate total costs including integration and human correction effort.

Deepgram’s reference guide compares its Nova-2 and Nova-3 speech-to-text models, framing Nova-2 as a lower-cost, high-throughput option for primarily English batch transcription and Nova-3 as a more capable but more expensive choice for real-time, noisy, multilingual, and domain-specific audio. It reports that Nova-2 offers fast batch processing, approximately 8.4% batch word error rate, and batch pricing of $0.0043 per minute, while Nova-3 supports live code-switching across 10 languages, keyterm prompting for up to 100 specialized terms, medical variants, real-time PII redaction, and reported word error rates of 5.26% for batch and 6.84% for streaming. The guide recommends selecting models based on required accuracy, latency, language coverage, customization, and volume-based costs, and suggests routing clean English segments to Nova-2 while directing noisy, multilingual, or high-stakes segments to Nova-3. It also outlines production practices including retries, latency monitoring, fallback logic, formatting and redaction settings, keyterm caching, and testing both models on representative audio before deployment.

Google Cloud experienced a widespread outage on June 12, 2025, after a policy update activated a bug in dormant code associated with regional Spanner tables used by Service Control, disrupting services including Google Speech-to-Text. The interruption affected organizations dependent on Google’s speech recognition APIs, preventing workloads such as call transcription, captioning, and real-time audio processing. Deepgram states that customers using its platform as a backup or replacement were able to continue operating with minimal disruption, attributing this outcome to a multi-cloud architecture, geographically distributed infrastructure, and redundancy across API, load-balancing, model-serving, and storage systems. The company reports increased interest from customers seeking alternatives after the outage and says its API allowed some Google Speech-to-Text users to migrate with limited code changes. The incident is presented as evidence that cloud reliability depends not only on the reputation of a provider but also on architectural choices such as provider diversity and independent failover capacity.

Deepgram announced the general availability of its Voice Agent API, a unified voice-to-voice platform that combines speech-to-text, text-to-speech, LLM orchestration, and conversational logic to help developers build real-time, context-aware voice agents. The API supports Deepgram’s Nova-3 speech recognition and Aura-2 speech synthesis models while allowing customers to bring their own LLMs or TTS systems, and it includes native handling for streaming audio, turn-taking, interruptions, session state, and dynamic behavior changes. Designed to avoid the tradeoff between inflexible low-code platforms and complex do-it-yourself integrations, it can be deployed in cloud, VPC, or on-premises environments for enterprise security and compliance needs. Deepgram says customers including Aircall, Jack in the Box, StreamIt, and OpenPhone use its technology to reduce costs, shorten wait times, and improve customer interactions. The company reported that its Voice Agent API ranked highest in its VAQI benchmark, surpassing OpenAI by 6.4% and ElevenLabs by 29.3%, and offers consolidated pricing of $4.50 per hour for users of its full stack, with discounts for those using external models.

Deepgram has made its Voice Agent API generally available, offering a unified streaming interface that combines Nova-3 speech-to-text, LLM orchestration, and Aura-2 text-to-speech while also allowing developers to supply their own language and speech models. The platform is designed to reduce the engineering complexity and latency associated with connecting separate voice AI services, with built-in capabilities for turn-taking, interruption handling, asynchronous function calls, live prompt updates, voice switching, and event-driven session control. Deepgram reports that its API ranked first in its Voice Agent Quality Index benchmark, which evaluated latency, interruption rates, and response coverage across 121 conversations, though the benchmark was developed and conducted by Deepgram. It supports managed cloud, private VPC, and on-premises deployments, including options intended for compliance-sensitive organizations, and provides Python and JavaScript SDKs plus an API playground. Priced at $4.50 per hour for its integrated stack, Deepgram says the service is less expensive than comparable offerings from ElevenLabs and OpenAI, while offering reduced pricing for customers that bring their own models.

Deepgram introduces the Voice-Agent Quality Index (VAQI), a 0-to-100 metric intended to quantify conversational voice-agent performance beyond standalone transcription accuracy or latency benchmarks. VAQI combines interruptions and missed response windows, each weighted at 40%, with latency weighted at 20%, using per-conversation normalization and logarithmic latency scaling to account for varying audio difficulty and extreme delays. The company evaluated providers including Deepgram, OpenAI, ElevenLabs, and Azure in repeated food-ordering simulations featuring pauses, self-corrections, background noise, and limited speaking opportunities, with synchronized audio streams and full-stack timing measurements. The results suggest that low latency alone is insufficient, as agents must also avoid talking over users and respond reliably when it is their turn; delays exceeding roughly three seconds substantially reduce perceived quality. Deepgram reports that its system achieved the highest internal VAQI score, above 70, while presenting the index as an ongoing benchmarking tool for engineering, procurement, and product improvement rather than a definitive measure of voice-agent quality.

An AI medical scribe can reduce clinicians’ documentation burden by combining Deepgram’s Nova-3 Medical speech-to-text model with OpenAI’s GPT-4.1 to convert doctor-patient conversations into structured SOAP notes for electronic health records. The tutorial explains why medical-specialized transcription and speaker diarization are important for accurately capturing clinical terminology and distinguishing participants, then demonstrates two implementation approaches in Python: processing prerecorded audio files through Deepgram’s API and streaming microphone audio live through Deepgram’s WebSocket endpoint. In both workflows, the transcript is sent to an LLM using a strict prompt that organizes available information into Subjective, Objective, Assessment, and Plan sections while avoiding unsupported assumptions. The prerecorded system reads an audio file, transcribes it, and writes a generated note to disk, while the real-time version uses PyAudio, asynchronous queues, and WebSockets to collect live speech, display finalized transcript segments, and create a SOAP note when recording ends. The article presents these tools as a way to let physicians concentrate more fully on patient interaction, while noting implementation details, required API keys, dependencies, sample testing methods, and a linked code repository.

Deepgram has introduced Nova-3 Medical Streaming, a real-time version of its healthcare-focused speech-to-text model that aims to combine clinical-grade accuracy with ultra-low latency for applications such as telemedicine, bedside documentation, ambient scribing, and triage. The company reports that the model reduces word error rates by 11% compared with Nova-3 General streaming and 30% compared with Nova-2 Medical streaming, while achieving a 2.7-fold improvement in keyword recall over Nova-3 General streaming. Designed for challenging clinical audio involving multiple speakers, distant microphones, background equipment noise, and specialized terminology, it supports custom prompting for up to 100 medical terms and uses HIPAA-compliant security measures. The model is available through Deepgram’s existing streaming API across hosted regions, allowing current users to upgrade without additional integration work, while self-hosted options are planned for the future.

A tutorial explains how to build a React-based real-time transcription application using Deepgram’s low-latency Nova-3 speech-to-text model and its Live Audio WebSocket API. The application captures microphone access through the MediaDevices API, divides audio into roughly 250-millisecond chunks with MediaRecorder, streams those chunks over a persistent WebSocket connection, and displays incoming transcription results dynamically in the interface. It contrasts Deepgram’s prerecorded and live transcription endpoints, describes response data such as transcripts, confidence scores, and word timestamps, and walks through creating a Vite React project with state for the transcript and recording status plus references for the microphone stream, recorder, and socket. A single toggle button starts or stops recording, while cleanup logic stops tracks, closes the socket, and ends the recorder; the resulting foundation can be extended for interviews, meetings, podcasts, accessibility tools, live captions, voice agents, and multilingual audio applications.