Deepgram Alternatives: 10 Speech-to-Text APIs Compared

The best Deepgram alternative depends on what you're transcribing. For recorded calls and meetings where accuracy matters most, try AssemblyAI. For audio in many languages that switch mid-sentence, try Gladia or ElevenLabs Scribe. If you're already deep in one cloud, use Google Cloud Speech-to-Text, Amazon Transcribe or Azure Speech. If audio can't leave your servers, look at Speechmatics on-prem or self-hosted Whisper. If you want zero per-minute bills, run faster-whisper yourself.

Deepgram is still a strong product, so this isn't a "get out now" list. It's a guide to the real reasons teams compare speech-to-text APIs, and which one fits each reason.

Quick Comparison

What Deepgram Does Well

Start with the bar you're comparing against. Deepgram's models overview lists three current speech-to-text models:

• Flux STT, a streaming model with turn detection built in, made for voice agents that need to know when someone has finished speaking.
• Nova-3, the general-purpose model for meetings, captions, noisy audio and multilingual audio, in batch or streaming.
• Nova-2, for languages Nova-3 doesn't cover yet and for filler-word detection.

The Deepgram pricing page says speech-to-text is billed per second with no block rounding. It also sells text-to-speech (Aura-2 and the new Flux TTS), a Voice Agent API that bundles speech-to-text, a language model and text-to-speech under one rate, and self-hosted containers on Enterprise.

So people rarely leave Deepgram because it's bad. They look around for one of these reasons:

• Accuracy on their own audio. Benchmarks use someone else's recordings. A model that wins on podcasts can lose on accented call-center audio.
• Language coverage. Flux multilingual covers ten languages. If your users speak something else in real time, Nova-3 or another vendor has to fill the gap.
• One cloud bill. A company standard on AWS, Google Cloud or Azure can outweigh a few points of word error rate.
• Data residency. Some audio can't leave a building, and Deepgram's self-hosting sits on the Enterprise plan.
• Cost at volume. Add-ons like diarization and redaction are billed on top, so the real per-minute cost depends on your feature mix.

The Best Deepgram Alternatives

AssemblyAI: the closest like-for-like swap

Best for: recorded meetings, calls and podcasts, plus voice agents that need high accuracy.

AssemblyAI's models page recommends Universal-3.5 Pro for recorded audio. It supports 18 languages with code switching and prompting, and falls back to Universal-2 for anything outside that set, for 99 languages in total. For streaming, Universal-3.6 Pro Realtime covers 32 languages with code switching, and AssemblyAI lists it at under 300ms latency. Universal-Streaming English and Multilingual are cheaper streaming options.

Medical Mode and a beta Veterinary Mode are add-ons that fix drug names and clinical terms. AssemblyAI now sells its own text-to-speech model too, so you can build a whole voice agent on one API key, the same pitch Deepgram makes.

The trade-off: pricing is per hour of audio, and features like prompting and diarization are billed as add-ons. There's no self-hosted option.

OpenAI: if you already use the OpenAI API

Best for: teams already calling OpenAI who want one vendor.

OpenAI changed its lineup this year. The speech-to-text guide now says to start with gpt-transcribe for recorded audio and gpt-live-transcribe for live audio. It also says whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize and the translations endpoint are deprecated and shut down on February 26, 2027.

That matters if you're picking OpenAI for a new build. The guide's examples for speaker labels, word timestamps and subtitle formats still use the deprecated models, so check each feature you need against the current docs before you plan around it. Uploaded files are capped at 25 MB, so long recordings need compressing or splitting.

The trade-off: it's a general AI platform, not a speech specialist. You get fewer speech-specific controls than Deepgram or AssemblyAI.

Google Cloud Speech-to-Text: for Google Cloud teams

Best for: apps already running on Google Cloud.

Chirp 3 is Google's latest multilingual speech model. Google says it's more accurate and faster than earlier Chirp models, and it adds diarization and automatic language detection. It's only available on the Speech-to-Text V2 API, through streaming, short sync requests and batch recognition.

Google bills on audio processed, measured in one-second increments. If your storage, IAM and billing already live in Google Cloud, that integration is often the deciding factor.

The trade-off: you're on the V2 API and Google's request model, which takes more setup than a single REST call to Deepgram.

Amazon Transcribe: for AWS and contact centers

Best for: teams on AWS, especially for call audio.

Amazon Transcribe runs on what AWS calls a multi-billion parameter speech foundation model for streaming and recorded speech. AWS lists 100+ languages with custom vocabulary, language identification, speaker diarization, word-level confidence scores, vocabulary filters, redaction of sensitive information and custom language models. For contact centers, Transcribe Call Analytics adds sentiment, call categories and generative AI summaries.

Usage is billed in one-second increments with no minimum, according to the AWS pricing page. Some advanced features carry extra charges.

The trade-off: it's built for AWS. If your stack isn't there, the IAM and S3 plumbing is overhead you don't get with Deepgram.

Azure Speech: for Microsoft shops and edge devices

Best for: companies on Azure, and products that need speech at the edge.

Microsoft now calls it Azure Speech in Foundry Tools. It offers real-time transcription, a MAI-Transcribe Streaming API for low-latency streams, fast transcription for files, and Azure LLM Speech, an LLM-powered model on the fast transcription endpoint. You can add words to the base vocabulary or build custom models.

The standout is deployment. Microsoft says you can run Azure Speech in the cloud or at the edge in containers, which Deepgram only offers on Enterprise.

The trade-off: the product sprawls across speech, translation and voice, with lots of model and endpoint choices. Expect to spend time in the docs.

Speechmatics: when audio can't leave your servers

Best for: regulated industries, broadcasters and on-device apps.

Speechmatics' deployment docs cover a managed cloud plus on-prem options: containers, GPU containers, Kubernetes and a virtual appliance. Its site also features on-device transcription, with Adobe Premiere and Stenograph as case studies. It offers realtime and batch transcription, an Agent STT mode and integrations with LiveKit, Pipecat, Vapi and Twilio.

The pricing page bills per hour of audio, with volume discounts.

The trade-off: on-prem means you run it. Budget GPU capacity and operations time.

Gladia: for multilingual and code-switching audio

Best for: products whose users switch languages mid-conversation.

Gladia's models page lists two models. Solaria-1 is the default, with 100+ languages, async and live modes, and code switching. Solaria-3 targets European business audio in English, French, German, Spanish and Italian, but runs on recorded audio only and has no code switching.

Gladia bills per hour of audio, with diarization, custom vocabulary and enhanced punctuation included on the self-serve async plan, and it has migration guides written for teams coming from Deepgram.

The trade-off: pick the model carefully. The most accurate one doesn't stream and doesn't code-switch.

ElevenLabs Scribe: for multilingual accuracy and diarization

Best for: teams that also use ElevenLabs voices, or need 90+ languages.

The ElevenLabs speech-to-text docs say to use Scribe v3 for recorded audio and Scribe v3 Live for real time, across 90+ languages. Features include keyterm prompting, realtime diarization, turn prediction, entity detection and mixed-language audio. The API pricing page lists Scribe v3 per hour of audio and says it has 27% fewer errors than Scribe v2, a vendor figure.

If you're comparing voice stacks more broadly, our ElevenLabs alternatives guide covers the text-to-speech side.

The trade-off: ElevenLabs is mainly a voice-generation company. Speech-to-text is one product among many, not its core.

Rev AI: for English accuracy from a transcription company

Best for: media and legal teams that care most about English accuracy.

Rev AI sells async and streaming speech-to-text across 57+ languages, with forced alignment for word-level timestamps and options to deploy in the cloud or on-prem. Its pricing page lists Reverb transcription per hour, rounded to the nearest second with a 15-second minimum, and also resells Whisper Large.

Rev also published Reverb, Apache-2.0 inference code for its ASR and diarization models. The repo notes that licence covers the code only. The model weights have their own licences on Hugging Face, so check those before commercial use.

The trade-off: fewer voice-agent features than Deepgram, Flux or AssemblyAI's streaming models.

Whisper and faster-whisper: the open-source option

Best for: offline transcription, privacy-sensitive audio and teams with spare GPUs.

OpenAI's Whisper repo is MIT-licensed and still ships model sizes from tiny to large, plus a turbo model the README lists at about 8x the speed of large. The hosted whisper-1 API is being retired, but the open-source weights aren't affected.

faster-whisper reimplements Whisper on CTranslate2. Its README says it's up to 4 times faster than openai/whisper at the same accuracy, using less memory, with 8-bit quantization on CPU and GPU.

The trade-off: you build everything Deepgram gives you out of the box: streaming, diarization, endpointing, scaling and monitoring. Whisper is a model, not a service.

Where Deepgram Is Still Stronger

Stay on Deepgram if:

• You're building a voice agent. Flux's built-in turn detection solves the "when do I stop listening?" problem that most pipelines patch with a separate voice activity model.
• You want one vendor for the whole voice loop. Speech-to-text, Aura-2 and Flux TTS, and a Voice Agent API sit under one account and one billing page.
• Per-second billing matters. Short clips aren't rounded up to a block, so bursty traffic stays cheap.
• You need self-hosting later. Deepgram's containers run in your VPC or on-prem with NVIDIA GPUs, so you can start in the cloud and move.
• Nova-3 already handles your languages. Its multilingual and monolingual coverage is broad, and self-serve keyterm prompting tunes vocabulary without retraining.

How to Choose a Deepgram Alternative

Before you switch, do three things:

1. Test on your own audio. Take 30 to 60 minutes of real recordings, including your worst ones, and compare word error rate and speaker labels by hand.
2. Price your real feature mix. Diarization, redaction and keyterms are add-ons at several vendors. Compare totals, not base rates.
3. Check latency end to end. For voice agents, measure time to final transcript and turn detection in your own pipeline, not the vendor's demo.

If you're shopping for a finished voice-agent platform instead of raw speech-to-text, our Vapi alternatives guide compares those. And if you just want meeting notes rather than an API, see the Otter.ai alternatives list.

Where Taku Fits

Taku isn't a speech-to-text API, and it doesn't replace Deepgram or anything on this list. It helps with a different problem: someone has already built a working transcription or voice tool on one of these APIs, and you'd rather use it than wire up your own.

That's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.

FAQ

Who are Deepgram's main competitors?

AssemblyAI, Speechmatics, Gladia, Rev AI and ElevenLabs Scribe among speech specialists, plus Google Cloud Speech-to-Text, Amazon Transcribe, Azure Speech and OpenAI among the big platforms. Open-source Whisper competes on cost.

What is the best open source Deepgram alternative?

Whisper, run through faster-whisper for speed. Both are MIT-licensed. Rev's Reverb is another option, but its Apache-2.0 licence covers the code, and the model weights carry their own licences.

What is the cheapest speech to text API?

It depends on volume and features. Most vendors bill per minute or per hour of audio and charge extra for diarization or redaction, so compare your total, not the headline rate. At high volume, self-hosted faster-whisper has no per-minute fee, but you pay for GPUs and engineering time.

Deepgram vs AssemblyAI: which is better?

Neither wins everywhere. Deepgram is stronger for voice agents, with Flux's built-in turn detection and per-second billing. AssemblyAI's Universal-3.5 Pro is built for accuracy on recorded audio, with prompting and code switching. Test both on your own recordings.

Deepgram vs Whisper: which should I use?

Use Deepgram if you want a managed service with streaming, diarization and scaling handled for you. Use open-source Whisper if you want offline transcription, full control over your data, and no usage fee. Note that OpenAI's hosted whisper-1 API shuts down on February 26, 2027.

Can I switch from Deepgram without rewriting my app?

Not quite. Every vendor has its own request format, streaming protocol and response shape. Wrap transcription behind one function in your code, and a later switch only touches that layer.