Deepgram vs ElevenLabs: Speech-to-Text Comparison

November 7, 2026

Deepgram and ElevenLabs both sell speech-to-text and text-to-speech, so “Deepgram vs ElevenLabs” can mean two different comparisons. Deepgram offers Nova speech recognition and Aura voices; ElevenLabs offers Scribe speech recognition and its voice generation models. This guide compares the speech-to-text side: Deepgram Nova-3 and ElevenLabs Scribe, both available through one Speechall API key. If you are choosing a voice for generated speech, the short section on text-to-speech below points you to each provider’s own documentation.

Neither model is the better choice for every recording. Accents, background noise, overlapping speakers, and domain vocabulary all change results. The practical way to decide is to send the same audio to both models and compare transcripts, latency, and cost on your own workload. The last section shows how to do that with two requests that differ only in the model ID.

Rates and capabilities below come from the Speechall model catalog as of November 7, 2026. They are Speechall credit usage, can change, and do not include credit purchase fees. Check each model page before you build a cost estimate.

Deepgram Nova-3 vs ElevenLabs Scribe at a glance

Nova-2 is included because it is the lower-cost Deepgram option with a longer language list in the catalog. For older recordings or price-sensitive batch work, it is worth adding to your test.

Two notes on the table. First, the language counts are what the Speechall catalog lists for each model. Scribe’s list covers the languages ElevenLabs publishes that Speechall accepts as a language code, and Nova-3 lists far fewer, so check the Scribe model page and the Nova-3 model page for your language. ElevenLabs also groups its languages by expected accuracy, and lower tiers can have much higher error rates. Second, the catalog lists Scribe v1. ElevenLabs also markets a Scribe v2 Realtime model for live transcription, which is not in the Speechall catalog today.

Where the two models differ in practice

Live audio. Through Speechall, Nova-3 and Nova-2 accept live audio over the streaming WebSocket endpoint. Scribe v1 is a file transcription model in the catalog. If you are building live captions, voice agents, or call transcription on Speechall today, start your shortlist with the Deepgram models.

Subtitles. Scribe returns both SRT and VTT through Speechall. The Deepgram models return VTT. If your video pipeline expects SRT files, Scribe saves a conversion step.

Long files. The catalog lists a 2-hour, 1 GB limit for Scribe and a 4-hour, 2 GB limit for Nova-2. Nova-3 has no limit listed in the catalog, so confirm with a representative file before relying on very long uploads.

Cost. For recorded files, Nova-3 is $0.312 and Scribe is $0.40 per audio hour on Speechall. At 100 hours a month, that is about $31.20 and $40.00 of credit usage. The cost guide covers retries, review time, and credit purchase fees, which often matter more than the rate difference.

Accuracy. Both companies publish their own benchmark results, and each tends to show its model favorably on its chosen datasets. Treat vendor benchmarks as a starting point, then measure word error rate on recordings that look like yours.

If you meant text-to-speech

Speechall does not offer text-to-speech today; its text-to-speech documentation describes it as a planned feature. To compare voices, go to each provider directly. Deepgram’s Aura voices are listed on the Deepgram pricing page, which bills text-to-speech per 1,000 characters of input text. ElevenLabs lists its voice models on its API pricing page. Check both pages for current rates.

When to choose each model

Choose Deepgram Nova-3 when:

• you need live transcription through Speechall’s WebSocket endpoint;
• your audio is in one of the languages listed on its model page;
• you want one model for both recorded files and live audio.

Choose ElevenLabs Scribe when:

• you transcribe recorded files and need SRT subtitles directly;
• your language appears in Scribe’s catalog list but not Nova-3’s;
• your tests show better transcripts on your audio, which justifies the higher hourly rate.

Choose Deepgram Nova-2 when cost matters more than having the newest Deepgram model, or when your recordings run longer than Scribe’s listed 2-hour limit.

When to integrate directly instead

Using both models through Speechall means one API key, one request format, and one credit balance. That helps while you are comparing models or when you expect to switch. Going directly to a provider can be the better choice when:

• you need a feature Speechall does not expose, such as ElevenLabs’ Scribe v2 Realtime model or a Deepgram feature outside the documented Speechall parameters;
• you have negotiated volume pricing or contract terms with one provider;
• your compliance review requires a direct data processing agreement with the provider;
• you have settled on one model and do not plan to switch.

The gateway vs direct integration guide walks through these tradeoffs in more detail.

Test both models on the same audio

Create an API key, then send the same file to both models. Only the model query parameter changes. The getting started guide covers authentication and the other upload methods.

AUDIO=meeting.wav

for MODEL in deepgram.nova-3 elevenlabs.scribe-v1; do
  curl -s -X POST "https://api.speechall.com/v1/transcribe?model=$MODEL&language=en&output_format=json&diarization=true" 
       -H "Authorization: Bearer $SPEECHALL_API_KEY" 
       -H "Content-Type: audio/wav" 
       --data-binary @"$AUDIO" 
       -o "transcript-$MODEL.json"
done

Set language=auto to let each model detect the language, or change output_format to srt or vtt to compare subtitle output where the model supports it. Then compare the two transcripts against a human-checked reference:

• word errors on names, numbers, and domain terms;
• speaker turns on overlapping speech;
• processing time for files of the length you expect;
• credit usage, shown in your console.

Run the test on several recordings, not one. A model that wins on a clean podcast can lose on a noisy phone call. Once you have a winner, keep the other model ID in your configuration: switching back on Speechall is a one-parameter change.

Browse every speech-to-text model with its rates and features in the Speechall model catalog.