The useful Deepgram vs Whisper decision starts with two requirements: whether speech arrives live, and who will run the speech model. A hosted speech service, an open-source model, and a retiring API create different engineering commitments. For developers choosing a transcription component, those differences determine the streaming contract, deployment work, and bill long before an accuracy score can settle the choice.
Deepgram vs Whisper at a glance
Deepgram Nova-3 fits a managed speech-to-text component for live audio or completed recordings. Self-run Whisper fits teams that need to own model execution and accept the infrastructure work. OpenAI's hosted whisper-1 is a separate, deprecated service with a migration deadline.
Here, speech-to-text (STT) means turning speech into a transcript. It does not include the complete runtime that listens, reasons, calls tools, and speaks as a voice agent.
| Choice | What it is | Responsibility boundary |
|---|---|---|
| Deepgram Nova-3 hosted API | Managed STT for streaming and pre-recorded audio | Deepgram operates model execution; your application handles the connection, transcript events, and downstream behavior |
| Open-source Whisper, run by you | Downloadable code and model weights, such as large-v3 | Your team runs the model and builds or adopts the service around it |
OpenAI whisper-1 | Hosted file-transcription API powered by Whisper V2 | OpenAI operates model execution; existing applications need migration before its scheduled shutdown |
The local Whisper path is not one fixed service. large-v3, turbo, an optimized inference engine, and a third-party host can produce different speed, output, and cost. A comparison result belongs to that exact configuration, rather than to every product carrying the Whisper name.
For a complete agent, the runtime is a separate buying decision. At Dasha, we provide a managed production platform for conversational AI products. We are relevant when your team wants to build and run agents, rather than replace a standalone transcription endpoint.
Streaming: live input, partial text and finished turns
“Streaming” can describe three different things:
- Live audio input: audio arrives continuously from a microphone or call.
- Incremental text output: transcript fragments arrive before processing finishes.
- Turn completion: the application receives a signal that it can respond without cutting the speaker off.
Support for one does not establish the other two.
Deepgram Nova-3 provides a live transcription service
Nova-3 accepts live audio through Deepgram's streaming connection and can return interim and finalized results. Completed recordings use POST /v1/listen, with either uploaded bytes or a remote audio URL.
With Nova, endpointing uses voice activity detection, which identifies speech and pauses. The configurable pause threshold can finalize text and set speech_final: true. That is a pause-based signal. Nova-3 does not include model-native turn detection; Deepgram's Flux is a separate model with that capability. The distinction is explicit in the model capability overview.
For a caller saying “Send it to…” and then pausing to recall an address, a final transcript segment and permission for the agent to speak are different events. An application that treats every finalized fragment as a completed turn can answer too early. Endpoint settings, transcript stability, and the agent's response policy all affect that behavior.
Whisper can support live systems through an added streaming layer
The reference Whisper transcribe() workflow reads a recording and processes it with sliding windows. Running the model alone does not provide a managed live-audio connection, reconnect handling, or a production event contract.
Streaming implementations can add buffering, repeated decoding, and rules for committing stable words. The independent Whisper-Streaming research demonstrates this approach using an agreement policy over successive outputs. That establishes that live Whisper systems are possible. It does not establish the latency of every Whisper deployment or parity with a hosted streaming API.
For a self-run system, the streaming wrapper is part of the product you operate. Its treatment of corrections, silence, context, and reconnections matters alongside the model.
OpenAI's current live transcription models are separate
whisper-1 does not support streamed output: its API ignores stream=true. It is a file-upload route, not a live microphone endpoint.
OpenAI recommends separate models: gpt-transcribe for completed recordings and gpt-live-transcribe for incoming live audio. A completed file can produce streamed text without becoming a live-input workflow. gpt-realtime-whisper is another distinct model available to existing integrations where supported; it is not the recommended starting point for a new integration.
OpenAI has deprecated whisper-1 and scheduled its API shutdown for February 26, 2027. The transcription retirement notice applies to that hosted service. It does not withdraw the separately released open-source Whisper code and weights.
Accuracy: compare the exact model and output contract
There is no defensible universal accuracy winner across Nova-3 and every Whisper configuration. A word error rate from one dataset cannot settle a different language, codec, microphone, or deployment.
Word error rate (WER) counts substitutions, deletions, and insertions relative to a reference transcript. It is useful, but it weights an incorrect account-number digit like any other word error. Product outcomes also depend on whether the transcript preserves names, amounts, dates, negation, and corrections.
A useful comparison separates four outputs:
| Output | What the result needs to establish |
|---|---|
| Final transcript | Error rates on the languages, accents, noise, and audio channels the product serves |
| Important entities | Correct names, numbers, identifiers, and domain terms |
| Live events | Stable partial text, usable finalization, and acceptable delay after speech ends |
| Required metadata | Compatible timestamps, speaker labels, and formatting |
Language coverage is also configuration-specific. Nova-3's hosted language=multi mode supports 55 languages and code-switching, meaning language changes within a conversation, while its monolingual configurations have their own language options. Whisper supports multilingual recognition, but its language-level results show substantial variation. A language count is an availability measure, not an accuracy score.
Speaker diarization, the assignment of speech to speakers, needs separate attention. Deepgram's diarize_model=latest setting currently resolves to v2 for batch and v1 for streaming. Streaming does not support v2. The self-run reference Whisper model needs a separate diarization layer if speaker labels are required; model inference alone is not a speaker-labeling service.
A credible latency result also names its endpoints. Time to the first changing word, time to stable text after speech ends, and time to the agent's first audible response measure different things. A file-processing throughput figure cannot answer a live turn-latency question.
Deployment: who owns the service around recognition?
Hosted Nova-3 removes inference operations
With Deepgram's hosted API, Deepgram manages the recognition infrastructure. Your team still owns audio capture, application credentials, connection recovery, event interpretation, data handling, and the product using the transcript.
Deepgram also offers a separately contracted commercial self-hosted deployment on customer infrastructure. The deployment responsibility split assigns infrastructure, monitoring, and backups to the customer. It is not equivalent to downloading MIT-licensed model weights.
Self-run Whisper gives control over model execution
Whisper's code and weights use the MIT License. The reference environment uses Python, PyTorch, and ffmpeg. Your deployment can keep inference in an environment you control, provided the surrounding storage, logging, and application paths remain inside that boundary too.
That ownership comes with concrete work:
- Packaging and pinning the model, inference engine, and dependencies.
- Provisioning compute, managing queues, and handling peak demand.
- Maintaining streaming and diarization layers where required.
- Monitoring failures, applying updates, and retaining a rollback path.
For offline transcription or a product with a firm local-execution requirement, that control can justify the work. For an irregular workload needing a service endpoint, idle capacity and operational ownership can outweigh a lower inference cost.
A third-party Whisper host changes this boundary again. The host operates its implementation and sets the billing meter, concurrency, and output contract. “Whisper-compatible” does not establish identical behavior to either the OpenAI API or your local deployment.
Cost: audio minutes and compute hours are different units
Deepgram's direct hosted price table lists these Nova-3 streaming rates. The Growth plan requires an annual prepaid commitment; its lower usage rate is not the same purchase as pay as you go.
| Nova-3 streaming item | Pay as you go | Growth |
|---|---|---|
| Monolingual transcription | $0.0077 per minute | $0.0065 per minute |
| Multilingual transcription | $0.0092 per minute | $0.0078 per minute |
| Speaker-diarization add-on | $0.0020 per minute | $0.0017 per minute |
Deepgram bills normal audio requests per second without block rounding. Multichannel processing is billed across the processed channels. Smart formatting is included. Pre-recorded transcription costs less, but the canonical direct API pricing page does not provide a numeric Nova-3 pre-recorded rate. These streaming figures do not establish a batch-specific price comparison. The request-usage API exposes per-request costs in USD. Deepgram's Amazon SageMaker Marketplace deployment has a separate rate card and must remain a separate budget.
OpenAI's deprecated whisper-1 model lists $0.006 per audio minute, equivalent to $0.36 per audio hour. That rate is relevant to an existing integration until shutdown. It is neither the cost of local Whisper nor the price of OpenAI's replacement models.
For new OpenAI integrations, gpt-transcribe file transcription costs $0.0045 per audio minute, while gpt-live-transcribe live transcription costs $0.017 per minute. These are separate file and live workflows, not a shared Whisper rate or a claim of equivalent output.
For 10,000 billable streaming audio minutes, the Nova-3 pay-as-you-go arithmetic is $77 for monolingual or $92 for multilingual, plus $20 if the listed diarization add-on applies to those minutes. The same volume of file transcription at the whisper-1 rate is $60. These are usage lines for different workflows, not complete application budgets or evidence that the outputs are interchangeable.
When can self-hosted Whisper cost less?
Self-hosting costs less when its fully loaded expense for sufficient capacity falls below the comparable hosted bill. No universal monthly-minute threshold proves that result.
A useful equation is:
self-host cost per accepted audio minute = (provisioned compute + storage + networking + operating work) / accepted audio minutes
“Accepted” excludes failed or unusable output. Provisioned compute includes paid idle time and spare capacity. A worker's audio-processing throughput determines whether it can meet the workload deadline at all.
A batch system that shuts down between jobs has a different cost shape from an always-ready live service. Live concurrency can require spare workers even when monthly audio volume is low. Faster offline processing can reduce compute hours, while stricter response deadlines can require more capacity.
Our Whisper pricing guide separates current OpenAI replacements, third-party hosting meters, and a worked self-hosting budget. Those deeper calculations belong to the deployment budget, rather than to an unsupported claim that free model weights make transcription free.
Which choice fits your product?
A managed STT component for live and recorded audio: Nova-3 provides those service workflows without your team running the model service. The remaining application work includes endpoint behavior, recovery, and transcript integration.
Local execution or control over the model environment: self-run Whisper provides downloadable code and weights. Its fit depends on accepting ownership of the inference service and any streaming or speaker-labeling layers.
An existing whisper-1 integration: the immediate issue is migration before February 26, 2027. The replacement needs to preserve the outputs the application consumes. A new model name alone does not guarantee compatible timestamps, subtitle formats, translation, or speaker labels.
If neither architecture meets the requirement, our Deepgram alternatives comparison covers the broader STT and voice-platform shortlist without treating those layers as interchangeable.
Where Dasha fits when the product is a voice agent
Recognition produces text. A voice agent also needs conversation execution, model responses, speech output, integrations, and channel operation.
At Dasha, we provide that managed runtime through a web application and REST APIs. Technical teams configure agents, prompts, models, tools, and webhooks, then deploy through phone or web integrations. Our Call Inspector exposes completed-call transcripts, recordings, model and tool interactions, and a conversation timeline.
Your team retains its business rules, data, downstream systems, telephony and compliance responsibilities, and production acceptance. We are not a drop-in Nova-3 or Whisper transcription endpoint, and this runtime choice does not imply support for every speech model.
If that is the layer your team wants managed, start with Dasha and build one narrow agent flow around your product's real integration.
