Deepgram vs. Whisper for phone bots: test the call, not the demo

Picking speech-to-text for a phone bot is not a contest between two names. It is a test of how each configuration handles your callers, your phone network and the action your bot must take next.

That distinction matters on PSTN calls. Narrowband audio, codecs, background noise, accents, overlapping speakers and dropped syllables all affect what reaches the recognizer. A transcript that looks acceptable to a person may still change a booking date, mishear a name or cause an intent classifier to choose the wrong path.

First, define what you mean by Deepgram and Whisper

Deepgram is a speech-recognition service. Whisper is a model family that can be used through different deployments. The comparison is therefore not necessarily service against service: hosting, model variant, decoding settings and the way a system marks the end of an utterance can all affect results.

Deepgram is a sensible candidate when you want a managed speech-recognition provider and can test its service against your calls. Whisper may suit teams that want to evaluate its model family in a deployment that fits their own infrastructure and operating requirements. Neither description predicts which one will perform better on a particular phone line.

Check the actual configuration before comparing. Record the model or service version, language settings, audio format, endpointing behavior and any custom vocabulary. If one system receives cleaner or differently processed audio, the test is not useful.

Build a PSTN test set that resembles production

Use real, permissioned calls or recordings that match your business: the same phone path, codec and typical caller environments. Include quiet calls, background television or traffic, different accents, fast speech, interruptions and cases where a caller corrects themselves. Keep a separate set of especially difficult examples so they do not disappear inside an average.

For each clip, create a human-checked reference transcript. Decide in advance how to treat punctuation, number formatting, disfluencies and partial words. Word error rate (WER) counts substitutions, deletions and insertions against that reference. It is useful for comparing transcripts, but its score changes with transcription conventions and can hide a critical error in a short phrase.

Report WER by condition as well as overall. A single blended score can conceal that one configuration works well in quiet audio but struggles with noisy calls. Also inspect errors that alter names, dates, amounts, negation or other details your bot uses to make decisions.

Measure delay and intent separately

For a live bot, transcription quality is only one part of the turn. Measure the time from the caller finishing a phrase to the transcript your application can act on. Keep partial transcript timing separate from final transcript timing, and use the same definition for both candidates. A system that appears faster because it finalizes incomplete phrases may create more corrections or wrong actions.

Then run both transcripts through the same intent-extraction and business rules. Score whether the bot identifies the caller’s request, captures required fields and takes the correct next step. Track cases where the transcript is wrong but the intent remains right, and cases where a small transcription error changes the action. This reveals whether additional recognition accuracy matters to your workflow.

For example, a bot may need to distinguish “Tuesday” from “Thursday” or “cancel” from “can’t cancel.” Overall WER does not tell you how often those specific mistakes happen. Keep a separate list of high-impact phrases and report their error rate.

Choose by evidence and operating fit

Prefer the configuration that performs reliably on your representative calls and meets your response-time needs. If results are close, weigh operational fit: where audio is processed, how the service is monitored, how changes are managed and whether your team can support the deployment. Do not infer a winner from a clean-room benchmark or a short demo.

For teams using Voicetta, the company lists Deepgram among its supported AI providers and says customers can use their own providers. That does not establish a Whisper integration or settle which recognizer will work best for a given call flow. Verify available configurations and test them with the actual audio path before relying on results.

The transcript also becomes part of the customer record and any later follow-up. Set rules for uncertain fields and make errors visible rather than silently treating every transcript as ground truth. The practical case for defined fields and uncertainty is covered in this guide to putting a contract around post-call data.

The short answer to “Deepgram vs. Whisper telephony?” is to compare named configurations on the same PSTN clips, then judge both transcript quality and the action the bot takes. If you cannot reproduce the test, you do not yet have a reliable basis for choosing.