Best transcription APIs for real-time voice applications
3 models · updated 2026-07-18
The verdict
Deepgram leads — All 3 models rank Deepgram the top pick.
As of 2026-07-18, ChatGPT, Claude, Gemini collectively rank Deepgram first for transcription apis for real-time voice applications on modelsagree.com.
Your vendor missing? Check any brand →
Combined ranking
- 1GPT #1Claude #1Gemini #1
Best overall for English-first voice agents: strong Nova-3-class recognition plus model-integrated turn detection, interruption/resumption events, streaming partials, and roughly 260ms median end-of-turn detection; its eager endpoint events can start LLM work before a turn is fully committed.
+ model takes & fixes− hide details
GPT Best overall for English-first voice agents: strong Nova-3-class recognition plus model-integrated turn detection, interruption/resumption events, streaming partials, and roughly 260ms median end-of-turn detection; its eager endpoint events can start LLM work before a turn is fully committed.
Claude Purpose-built for streaming from day one — Nova-3 delivers ~300ms end-to-end latency over a real WebSocket API with word-level timestamps, endpointing, and interim results that actually behave predictably; pricing (~$0.0077/min streaming) undercuts most rivals at scale, and self-hosted/VPC deployment exists for regulated shops. For the typical voice-agent builder it has the best latency-accuracy-cost balance, which earns #1.
Gemini Industry-leading low latency, particularly with the Flux model featuring model-native turn detection under 260ms, and exceptionally high accuracy for conversational voice agents.
Where it falls shortper GPT Flux’s language coverage is narrower than the strongest multilingual APIs, and extracting maximum responsiveness requires tuning turn thresholds and handling speculative responses.
per Claude Accuracy on heavily accented or far-field noisy audio still trails the best batch models, and its language coverage is thinner than Speechmatics or Google — English-first workloads shine, long-tail languages don't.
- 2GPT #2Claude #2Gemini #4
Near-tie for first on recognition quality, especially names, numbers, emails, and domain terms; combines roughly 150ms post-endpoint latency with semantic endpointing, dynamic keyterm prompting, immutable finals, straightforward WebSockets, and an unusually practical self-hosting option.
+ model takes & fixes− hide details
GPT Near-tie for first on recognition quality, especially names, numbers, emails, and domain terms; combines roughly 150ms post-endpoint latency with semantic endpointing, dynamic keyterm prompting, immutable finals, straightforward WebSockets, and an unusually practical self-hosting option.
Claude Universal-Streaming closed the gap with Deepgram on latency (~300ms immutable transcripts) while generally edging it on English accuracy, and its endpointing/turn-detection tuned for voice agents reduces the awkward-interruption problem that plagues LLM voice bots. Strong docs and per-second billing make it easy to adopt. Near-tie with Deepgram — Deepgram wins on price and deployment options, AssemblyAI on out-of-box turn handling.
Gemini Excellent developer experience with robust SDKs, offering sub-second latency alongside direct streaming integration with their audio intelligence suite for features like PII redaction.
Where it falls shortper GPT At about $0.45 per session-hour it costs materially more than value-oriented alternatives, while self-hosting requires a substantial commercial commitment.
per Claude Streaming is English-centric (multilingual streaming support lags well behind its batch offering), and there's no self-hosted option for data-residency-constrained teams.
- 3GPT #3Claude —Gemini #2
Near-tie with Deepgram on speed, achieving latency as low as 150ms using predictive transcription models, combined with highly aggressive pricing at 0.39 dollars per audio hour.
+ model takes & fixes− hide details
Gemini Near-tie with Deepgram on speed, achieving latency as low as 150ms using predictive transcription models, combined with highly aggressive pricing at 0.39 dollars per audio hour.
GPT The strongest default when multilingual reach matters: approximately 150ms partial latency, automatic language recognition across 90-plus languages, word timestamps, VAD, manual commit control, and native support for telephony audio.
Where it falls shortper GPT Turn-taking controls are less conversation-native than Flux’s, and plan-based concurrency and pricing can become awkward at production scale.
- 4GPT #4Claude #3Gemini #5
Best-in-class multilingual and accent robustness in real-time — 50+ languages with one model family, strong diarization, and flexible deployment (SaaS, container, on-prem), making it the default when your callers aren't American English speakers. Ursa models hold accuracy at low latency better than most.
+ model takes & fixes− hide details
Claude Best-in-class multilingual and accent robustness in real-time — 50+ languages with one model family, strong diarization, and flexible deployment (SaaS, container, on-prem), making it the default when your callers aren't American English speakers. Ursa models hold accuracy at low latency better than most.
GPT Excellent multilingual and accented-speech performance, mature partial/final transcript handling, strong customization, diarization, and cloud, on-premises, or appliance deployment make it a dependable choice for global or regulated applications.
Gemini Superb accuracy in noisy environments using the Ursa engine, with granular control over latency-accuracy trade-offs via a configurable delay parameter down to 0.7 seconds.
Where it falls shortper GPT Pricing and deployment terms are comparatively sales-led, and its developer experience is less frictionless for a small team seeking instant pay-as-you-go voice-agent deployment.
per Claude Costs meaningfully more than Deepgram/AssemblyAI and integration ergonomics (SDKs, examples, voice-agent tooling) trail the US developer-first vendors — it's NOT the cheapest or fastest path to a demo.
- 5GPT —Claude —Gemini #3
Delivers Whisper-level accuracy optimized for live streaming with a final transcript latency of 270-300ms, showing outstanding capabilities in handling real-time, mid-sentence language code-switching.
+ model takes & fixes− hide details
Gemini Delivers Whisper-level accuracy optimized for live streaming with a final transcript latency of 270-300ms, showing outstanding capabilities in handling real-time, mid-sentence language code-switching.
- 6GPT —Claude #4Gemini —
If you're already building the agent on OpenAI, transcription arrives inside the same Realtime session — one vendor, one WebSocket/WebRTC connection, with semantic VAD and strong accuracy from the audio-native model; simplest total architecture for speech-to-speech products.
+ model takes & fixes− hide details
Claude If you're already building the agent on OpenAI, transcription arrives inside the same Realtime session — one vendor, one WebSocket/WebRTC connection, with semantic VAD and strong accuracy from the audio-native model; simplest total architecture for speech-to-speech products.
Where it falls shortper Claude It's not a standalone STT tool — weaker controls (no word timestamps in streaming, limited formatting/diarization), occasional hallucinated transcript segments under noise, and pricing that beats dedicated STT vendors only if you're consuming the rest of the stack anyway.
- 7GPT #5Claude —Gemini —
Strong multilingual streaming recognition, broad regional and enterprise infrastructure, adaptation features, and dependable scaling make it valuable for teams already operating on Google Cloud; it nearly ties Speechmatics where cloud governance matters most.
+ model takes & fixes− hide details
GPT Strong multilingual streaming recognition, broad regional and enterprise infrastructure, adaptation features, and dependable scaling make it valuable for teams already operating on Google Cloud; it nearly ties Speechmatics where cloud governance matters most.
Where it falls shortper GPT It is comparatively expensive and lacks the purpose-built conversational endpointing and turn-event ergonomics of the leaders.
- 8GPT —Claude #5Gemini —
The strongest open/self-hosted route for real-time — Parakeet-family models top open ASR leaderboards, Riva gives production-grade streaming (gRPC, sub-second latency, GPU batching), and at sustained high volume it's dramatically cheaper per minute than any API while keeping audio entirely in your infra.
+ model takes & fixes− hide details
Claude The strongest open/self-hosted route for real-time — Parakeet-family models top open ASR leaderboards, Riva gives production-grade streaming (gRPC, sub-second latency, GPU batching), and at sustained high volume it's dramatically cheaper per minute than any API while keeping audio entirely in your infra.
Where it falls shortper Claude You operate GPUs and model serving yourself — for teams without infra muscle or with spiky low volume, the ops cost wipes out the savings; it's NOT for someone who wants a key and a WebSocket today.
Just missed the top 5
GPT Cartesia Ink-2 — excellent price, latency focus, and telephony orientation, but its STT platform has less production history and independently demonstrated breadth than the top five · Microsoft Azure AI Speech real-time transcription — broad language, compliance, and customization support, but weaker value and less voice-agent-native turn handling for the typical practitioner
Claude Gladia — good latency and multilingual streaming with Whisper-derived accuracy, but less proven at scale and thinner enterprise track record than the top four
Gemini OpenAI Realtime API — extremely cost-prohibitive for pure transcription since pricing is scaled for full audio-to-audio LLM pipelines · Google Cloud Speech-to-Text Chirp v2 — suffers from high endpointing latency that often exceeds 2 seconds unless customized with external VAD
By model
ChatGPT
- 1.Deepgram
- 2.AssemblyAI
- 3.ElevenLabs Scribe
- 4.Speechmatics
- 5.Google Cloud Speech-to-Text
Claude
- 1.Deepgram
- 2.AssemblyAI
- 3.Speechmatics
- 4.OpenAI Realtime API
- 5.NVIDIA Riva
Gemini
- 1.Deepgram
- 2.ElevenLabs Scribe
- 3.Gladia
- 4.AssemblyAI
- 5.Speechmatics
Common questions
What is the best transcription apis for real-time voice applications according to AI models?
Deepgram leads. All 3 models rank Deepgram the top pick. The current top 3: Deepgram, AssemblyAI, ElevenLabs Scribe. Ranked by asking ChatGPT, Claude, Gemini the same buying question and merging their top-5 picks, updated 2026-07-18. Source: modelsagree.com.
Which transcription apis for real-time voice applications did each AI model pick first?
ChatGPT: Deepgram. Claude: Deepgram. Gemini: Deepgram.
How is this transcription apis for real-time voice applications ranking made?
ChatGPT, Claude, Gemini are each asked the same buying question in a fresh session with no system steering. Their top-5 answers are merged (rank 1 = 5 pts … rank 5 = 1 pt) into the consensus ranking, re-polled weekly and tracked over time.
More on how polling works: full methodology →
This ranking moves
We re-poll all four models weekly. Get one short email when a #1 flips.
Cite this ranking
ModelsAgree, “Best transcription APIs for real-time voice applications” — merged ranking from ChatGPT, Claude, Gemini & Grok, polled 2026-07-18. https://modelsagree.com/best/best-transcription-apis-for-real-time-voice-applications (CC BY 4.0)
Tracked by ModelsAgree · rank 1 = 5 pts … rank 5 = 1 pt · re-polled weekly