Deepgram Reviews & Product Details

The speed is the standout feature. I'm using Nova-2 with dual WebSocket streams (mic + tab audio simultaneously) for a real-time voice coaching Chrome extension, and interim results come back fast enough to display a genuinely live rolling transcript. Word-level confidence scores, smart_format, vad_events, and utterance_end_ms all work exactly as documented and saved me a lot of custom logic. The $200 starter credit is also a great touch - it gave me real runway to prototype and validate the product before committing to a paid plan. Review collected by and hosted on G2.com.

Nothing major so far. The documentation could occasionally go deeper on edge cases around dual-stream WebSocket setups and silence/KeepAlive behavior, which required some trial and error to get right. But overall these are minor friction points in an otherwise smooth experience. Review collected by and hosted on G2.com.

The speech-to-text accuracy and punctuation are awesome compared with GCP and Azure STT. Pricing is also better than other services. Review collected by and hosted on G2.com.

Sometimes Nova 2 performs better than Nova 3, and Nova 3 still doesn’t support keywords. Also, the multi-language detection isn’t very accurate when you compare results across multiple streams. In my DilMesh app, I create a separate stream for each language that might be spoken. At our events we generally use Turkish and English, so the app creates two streams—one for Turkish and one for English—and then selects the final results based on the confidence response. Multi stream method works terrific when you compared with built-in multi language detection. Review collected by and hosted on G2.com.

The best thing about Deepgram is that it lets users try the platform’s STT and other modules first using trial credits, which helps them get a fair understanding of the product before deciding to buy it. Review collected by and hosted on G2.com.

So far, I haven’t figured out any issues that I would say I disliked. Review collected by and hosted on G2.com.

Real-time transcription speed and accuracy is exceptional. I integrated Deepgram into my Windows app StreamVox for live audio translation, and the API was straightforward to set up. The documentation is clear, WebSocket streaming works reliably, and latency is low enough for real-time subtitle overlays. The Nova model handles different accents well even with background noise. Review collected by and hosted on G2.com.

The dashboard could use more detailed usage analytics and graphs. It would be helpful to see per-request breakdown and latency stats in one place. Review collected by and hosted on G2.com.

We evaluated Deepgram for our medical education platform (PassBoards.ai (http://passboards.ai/)) and were impressed by the transcription accuracy with medical terminology — ophthalmology terms, drug names, anatomical structures. The API is clean and well-documented, latency is low, and their team (shoutout to Danny Kim) was responsive and genuinely helpful during our evaluation. For healthcare/medical applications where accuracy on specialized vocabulary matters, Deepgram stood out Review collected by and hosted on G2.com.

Pricing can be hard to forecast when you're scaling a startup. Would love to see more transparent volume-based tiers for small teams. Otherwise, no major complaints. Review collected by and hosted on G2.com.

I love that Deepgram's API is super easy to set up in just a couple of minutes. I never even go into the UI because the API alone is great. The instant transcription is quick, reliable, and quite accurate, which is essential for my freight brokerage, where we need to call truckers all the time. The fact that it works better than our phone system provider is a big plus. I honestly don't spend any time on it, which is the biggest compliment I could give since it just works. Also, the setup process couldn't be easier—I'd give it a nine out of ten. All I had to do was plug the API key into Claude Code and the whole thing set itself up. Their documentation is really good too. Review collected by and hosted on G2.com.

Honestly no complaints Review collected by and hosted on G2.com.

I find Deepgram works reliably, and the latency is quite low, which gives the feel of live functionality when transcribing speech. The live transcription feature is really fast, which I find great. For English, Deepgram performs very well and is mostly good for other languages too. Especially for Russian, the quality is superior compared to other tools. I appreciate that Deepgram's cloud is much faster than local alternatives like Whisper, and it's really GDPR compliant, offering EU servers which is a game changer for privacy. The initial setup was actually very easy due to clear documentation and a nice playground to try things before implementation. Review collected by and hosted on G2.com.

I tried using medical transcription, but it only works in English. For my work, it would be better and required to be in German. There's a special model designed specifically for medical language, mainly in English, and I would like to have it in German as well and in other languages also. Review collected by and hosted on G2.com.

I really like the Deepgram Flux model, especially its natural pause detection, which allows my AI agent to transcribe without sending an end event until I'm actually done talking. This feature gives it a more natural conversational ability rather than interrupting me while I'm still talking. I also like that Deepgram is really fast and really cheap. Review collected by and hosted on G2.com.

I mean, I guess it was kinda tricky to do the integration to make it super smooth with my agent, but I got it done, and it works well. Well, I'm using Cartesia for text to speech. And I was hoping to just use Deepgram Text to Speech, but I found that Cartesia has more natural voices. It was really easy except it needed a lot of tinkering to get the word timing working correctly. Review collected by and hosted on G2.com.

I think the SDK and WebSocket API are clean, which made setting up the streaming connection pretty straightforward. Running two separate WebSocket streams for my microphone and system audio worked well, and the connection was stable. The JSON message format also made sense, and the reconnection logic was easy to implement. From a developer experience standpoint, getting the basics up and running felt smooth. Nova three looked really promising with its impressive accuracy benchmarks and good latency numbers. Features like smart formatting and punctuation worked well out of the box. Deepgram has solid technology, and Nova three is a capable model with a good API design. Review collected by and hosted on G2.com.

The first big challenge was actually turn detection. Deepgram fires these utterance end events based on silence and the problem is it fires on every pause, not once per complete thought. So, in a real meeting, when someone is talking and they just pause for a breath or to think for half a second, it results in separate fragments. I spent a lot of time trying to tune utterance and milliseconds. It always felt like I was fighting the system instead of working with it. The second issue was real world accuracy. When I compared what Deepgram transcribed to what Loom captured, the differences were rough. Domain-specific words were getting hallucinated into completely wrong things, like 'contractor role' turning into 'podcast show.' And the key term prompting feature in Nova three, which could have helped, felt static, and I never got it properly integrated during development. The feature could be more accessible or documented better. The third thing was transcript mutability. Deepgram's streaming results are mutable, which makes sense technically, but for a real-time UI, it creates a flickering effect where text keeps changing. This made the experience feel unreliable from the user's perspective. Review collected by and hosted on G2.com.

I find Deepgram's UI to be intuitive, simple, and highly functional, making it easy for users to navigate and complete their tasks. The variety of models like Nova 2, 3, and Whisper allows me to fine-tune my usage based on language needs, which is very helpful for my Oral History Project. The sub-features within the model, such as diarization of speakers, redaction, sentiment analysis, and grammar, make it an excellent tool for working with ASR-related tasks. The fast transcription rates significantly improve time management by transcribing hours of audio in seconds, streamlining my workflow and enabling more interviews. Moreover, Deepgram's API is straightforward to integrate, creating an effortless experience for developing independent tools. The quality of transcriptions is greatly enhanced, and the ability to catalog based on efficiency, success rate, and accuracy is invaluable. Overall, Deepgram is not only reliable and cost-effective but is also a de facto choice for speech-to-text models. It's clearly a 10 for me, and I've already recommended it to colleagues because of the ongoing improvements and planned developments. Review collected by and hosted on G2.com.

The only issue is the inability to rename speakers and the absence of timestamps. I understand JSON files could solve this but it would save a lot of time. It takes me a while to extract all the details like timestamps and speaker changes from a JSON file. Furthermore, processing the word file results in a very large file size. Review collected by and hosted on G2.com.