Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

AssemblyAI vs Deepgram for batch transcription (2026): accuracy, throughput and pricing

AssemblyAI-STT-3.7 vs Deepgram Nova-3 for pre-recorded audio: call-center WER, speaker diarization, code switching, batch throughput, pricing and a migration snippet.

Abstract green cylinder illustration

Written by

Kelsey Foster

Published on

19 November 2026

If you're transcribing recorded audio at scale, such as call recordings, meetings, interviews or podcasts, the batch decision comes down to three questions. How accurate is the transcript on the audio you have? How many files can you push through at once? And what does an hour cost once you add the features your pipeline needs? This guide compares AssemblyAI-STT-3.7, AssemblyAI's flagship pre-recorded speech-to-text model, with Deepgram Nova-3 on each question, then shows the code change to move a batch job across.

AssemblyAI vs Deepgram for batch at a glance

Criterion AssemblyAI (AssemblyAI-STT-3.7) Deepgram (Nova-3)
Diarization error rate (6 English sets) 10.7% 17.4%
Code-switching error rate 6.2% 49.4%
Call-center WER (AppTek) 7.44% 8.59%
Base price, pre-recorded $0.21/hr $0.0043/min Pay as you go; $0.0036/min Growth (promo heading)
Speaker diarization +$0.02/hr Included for pre-recorded
Default parallel jobs (paid self-serve) 200+, queued beyond the limit 50 concurrent requests
Languages 32 natively code-switched; 99 with Universal-2 Monolingual and multilingual Nova-3 options, priced separately
Free credit $50 $200

AssemblyAI figures come from the AssemblyAI-STT-3.7 research post and the AssemblyAI docs. Deepgram figures come from Deepgram's pricing page and rate limits, checked November 19, 2026.

Why batch transcription needs a different scorecard

Batch work has different priorities from real-time transcription. Nobody is waiting on a sub-second response. You're working through a queue of files, and the transcript feeds something downstream: analytics, summaries, compliance records, search or an LLM. Three things decide whether that pipeline works.

  • Accuracy on recorded speech. Phone audio, accents, crosstalk and speakers who switch languages are where batch transcripts break.
  • Speaker attribution. Most batch audio has more than one speaker. A sentence credited to the wrong person corrupts every summary and QA score built on it.
  • Throughput and cost at volume. A backlog of 100,000 calls turns per-file limits and per-hour add-on fees into the real numbers.

Which is more accurate on recorded audio?

AssemblyAI-STT-3.7 posts a lower speaker diarization error rate than Deepgram Nova-3, a far lower error rate on code-switched speech, and a lower word error rate on public call-center audio. Every figure below comes from the launch of AssemblyAI-STT-3.7 and its research post, with each vendor's production pre-recorded API on default settings.

Speaker diarization

On English conversations, AssemblyAI-STT-3.7 has the lowest speaker diarization error rate of the pre-recorded APIs measured: 10.7%, averaged over six datasets. Deepgram Nova-3 is at 17.4%.

Provider Diarization error rate (lower is better)
AssemblyAI-STT-3.7 10.7%
ElevenLabs Scribe v3 12.0%
Speechmatics Enhanced 16.2%
Gladia Solaria-1 16.7%
AWS Transcribe 16.9%
Deepgram Nova-3 17.4%
Soniox stt-async-v5 24.2%

Compared with Universal-3.5 Pro at its launch, AssemblyAI's new diarization model cuts cpWER by 20%, DER by 36% and speaker-count errors by 63%, and the diarization step runs 3–4× faster. If you already set speaker_labels: true, you get it with no code change.

Code-switched audio

Code switching is where the field spreads out most. On audio that mixes English with German, Spanish, French, Italian or Portuguese, AssemblyAI-STT-3.7 has the lowest average error rate, 6.2%. Deepgram Nova-3 averages 49.4% on the same sets.

Provider Average code-switching WER (lower is better)
AssemblyAI-STT-3.7 6.2%
ElevenLabs Scribe v3 6.8%
Soniox stt-async-v5 10.2%
Azure Batch Transcription 17.4%
AWS Transcribe 22.9%
OpenAI GPT-4o Transcribe 44.7%
Deepgram Nova-3 49.4%

For a batch pipeline that ingests calls from multilingual customers, that gap shows up as transcripts that are usable or not. AssemblyAI-STT-3.7 code-switches natively across its 32 languages, so there's no separate multilingual model to pick.

Call-center word error rate

AppTek's call-center dialogues are a public dataset of English customer-service calls across 14 accents, scored with AppTek's own script. On a 526-recording subset (about 91 hours), AssemblyAI-STT-3.7 is ahead of Deepgram Nova-3, Gemini 3.5 Transcribe, Soniox, Speechmatics and OpenAI GPT-4o Transcribe.

Provider Word error rate (lower is better)
AssemblyAI-STT-3.7 (defaults) 7.44%
Gemini 3.5 Transcribe 7.52%
Soniox stt-async-v6 7.89%
Deepgram Nova-3 8.59%
Speechmatics Enhanced 8.76%
OpenAI GPT-4o Transcribe 10.82%

The dataset's references spell out every number, so AssemblyAI also ran it with format_text off, which returns spoken-form text. With that setting the error rate is 7.05%.

Run Your Own Batch Files

Upload a few of your hardest recordings to the playground and compare AssemblyAI-STT-3.7 transcripts against what you get today. No code required.

Try playground

How much batch throughput do you get?

Throughput is set by how many files you can have in flight and what happens when you exceed that.

AssemblyAI. Paid accounts get 200+ parallel transcriptions by default (5 on free accounts). Jobs beyond the limit aren't rejected: they go into a queue and run in first-in, first-out order as earlier jobs finish. Each request takes files up to 5 GB and up to 10 hours long, and AssemblyAI raises rate limits on request at no extra cost. See the pre-recorded rate limits in the docs.

Deepgram. Nova-3 pre-recorded requests are limited to 50 concurrent requests on both the Pay as you go and Growth plans. Enterprise starts at 200. Limits apply per project.

For a nightly backlog, a 200-job default against a 50-request default is a fourfold difference in files in flight before you talk to sales. Use webhooks rather than tight polling loops either way. AssemblyAI caps total API requests at 20,000 per 5 minutes, which a fleet polling every few seconds can hit.

Pricing for batch transcription

Both vendors bill pre-recorded audio by duration, and both price some features as add-ons. The table below lists per-hour rates for the pieces a batch pipeline usually needs.

Item AssemblyAI Deepgram Pay as you go Deepgram Growth ($4,000+/yr prepaid)
Base model, English $0.21/hr (AssemblyAI-STT-3.7) $0.0043/min (about $0.26/hr) $0.0036/min (about $0.22/hr)
Base model, multilingual $0.21/hr, same model $0.0052/min (about $0.31/hr) $0.0043/min (about $0.26/hr)
Speaker diarization +$0.02/hr Included Included
Keyterm prompting +$0.05/hr (up to 1,000 terms) +$0.0013/min (about $0.08/hr) +$0.0011/min (about $0.07/hr)
Entity detection +$0.08/hr +$0.0017/min (about $0.10/hr) +$0.0014/min (about $0.08/hr)
Redaction +$0.08/hr (PII text); +$0.05/hr (PII audio) +$0.0020/min (about $0.12/hr) +$0.0017/min (about $0.10/hr)

A few details change the math for real workloads.

  • Diarization. Deepgram includes diarization in its pre-recorded price. AssemblyAI charges $0.02/hr for it, so English batch with speaker labels runs $0.23/hr on AssemblyAI-STT-3.7.
  • Plans. Deepgram's Growth rates require an annual prepayment of $4,000 or more, with a 10% premium on overage beyond the prepaid credits. AssemblyAI's rates are pay-as-you-go with volume pricing through sales.
  • Promo labels. Deepgram lists its pre-recorded Nova-3 rates under a "Promo pricing" heading, so check the page before you budget.
  • Free credit. New AssemblyAI accounts get $50 in free credits for the speech APIs, with no credit card required. Deepgram offers $200 in non-expiring credit.
  • Multilingual. Deepgram prices Nova-3 multilingual separately ($0.0052/min Pay as you go). AssemblyAI-STT-3.7 is one price across its 32 languages, and Universal-2 covers the rest at $0.15/hr.

The full rate card is on the AssemblyAI pricing page. For a broader view across vendors, see our speech-to-text API pricing breakdown.

Price Your Backlog on Real Audio

Get $50 in free credits, point AssemblyAI-STT-3.7 at a sample of your batch audio, and see accuracy and cost per hour on your own files.

Sign up free

How to migrate a batch job from Deepgram

For pre-recorded audio, the request shapes are close. Here's a Nova-3 request against Deepgram's REST API with smart formatting and diarization:

import requests

resp = requests.post(
    "https://api.deepgram.com/v1/listen",
    params={"model": "nova-3", "smart_format": "true", "diarize": "true"},
    headers={"Authorization": "Token YOUR_DEEPGRAM_KEY"},
    json={"url": "https://example.com/call.mp3"},
)
print(resp.json()["results"]["channels"][0]["alternatives"][0]["transcript"])

And the same job on AssemblyAI-STT-3.7 with the Python SDK, which submits the file and polls until the transcript is ready:

import assemblyai as aai

aai.settings.api_key = "YOUR_ASSEMBLYAI_KEY"

config = aai.TranscriptionConfig(
    speech_models=["assemblyai-stt-3-7", "universal-2"],
    language_detection=True,
    speaker_labels=True,
    disfluencies=False,  # 3.7 keeps fillers by default; False gives clean text
)

transcript = aai.Transcriber().transcribe("https://example.com/call.mp3", config)

for utterance in transcript.utterances:
    print(f"Speaker {utterance.speaker}: {utterance.text}")

Four things to know when you switch:

  1. Formatting is on by default. format_text defaults to true, so there's no smart_format equivalent to set. Set it to false only if you want spoken-form text.
  2. Disfluencies are on by default. AssemblyAI-STT-3.7 keeps fillers, repetitions and false starts unless you set disfluencies to false. That's a change from Universal-3.5 Pro.
  3. Fallback is explicit. With language_detection on, list both models, ["assemblyai-stt-3-7", "universal-2"], so audio in a language outside 3.7's 32 falls back to Universal-2 for 99-language coverage.
  4. Use webhooks at volume. For large queues, pass a webhook URL instead of polling each job.

The Deepgram-to-AssemblyAI migration guide in the docs maps the remaining parameters one by one.

When AssemblyAI is the right call for batch

Choose AssemblyAI-STT-3.7 when your recorded audio feeds something that depends on getting words and speakers right:

  • Contact-center analytics and QA, where call-center WER and speaker attribution decide whether agent scores are correct.
  • Meeting and interview transcripts, where diarization errors turn into misattributed action items.
  • Multilingual customer audio, where speakers mix English with another language mid-call.
  • Large backlogs, where 200+ parallel jobs by default with queueing beyond that keeps a nightly run moving.

Deepgram's limits for batch are concrete: 50 concurrent pre-recorded requests on self-serve plans, a 17.4% diarization error rate and 49.4% on code-switched audio in AssemblyAI's published runs, and pre-recorded rates listed under a promo heading. For a view of how the two compare beyond batch, including streaming and voice agents, see AssemblyAI vs Deepgram, and for the wider field, our list of Deepgram alternatives.

The verdict

For batch transcription, AssemblyAI-STT-3.7 is ahead of Deepgram Nova-3 on the published measures that matter for recorded audio: diarization error rate (10.7%, the lowest of the pre-recorded APIs measured, vs 17.4%), code-switching error rate (6.2% vs 49.4%) and call-center WER (7.44% vs 8.59%). It starts at 200+ parallel jobs with queueing, costs $0.21/hr across 32 languages, and moving a job across is a short config change. The best confirmation is your own audio: run a sample of last month's files through both and compare the transcripts, speaker labels and cost per hour. If your errors cost money downstream, the true cost of inaccurate transcription is worth reading before you decide.

Start Your Batch Migration

Sign up free, swap in AssemblyAI-STT-3.7, and run your next batch with speaker labels and native code switching.

Sign up free

Frequently Asked Questions (FAQs)

Is AssemblyAI or Deepgram more accurate for batch transcription?

On AssemblyAI's published November 2026 runs, AssemblyAI-STT-3.7 is ahead of Deepgram Nova-3 on recorded audio. It posts 7.44% WER on AppTek's public call-center dataset against 8.59% for Nova-3, a 10.7% diarization error rate against 17.4%, and 6.2% on code-switched audio against 49.4%. Each vendor was run through its production pre-recorded API with default settings.

How much does Deepgram batch transcription cost compared with AssemblyAI?

Deepgram lists Nova-3 pre-recorded at $0.0043/min on Pay as you go (about $0.26/hr) and $0.0036/min on Growth (about $0.22/hr), which requires a $4,000+ annual prepayment. Diarization is included. AssemblyAI-STT-3.7 is $0.21/hr, plus $0.02/hr for speaker diarization. Deepgram labels its current rates as promo pricing, so check its page before budgeting.

How many files can I transcribe at once?

AssemblyAI paid accounts run 200+ pre-recorded transcriptions in parallel by default, and jobs beyond that are queued and processed in order rather than rejected. Deepgram's rate-limit page lists 50 concurrent pre-recorded requests on Pay as you go and Growth, and 200 to start on Enterprise. Both vendors raise limits through their sales teams.

Does AssemblyAI support diarization for batch transcription?

Yes. Set speaker_labels to true on any pre-recorded request. AssemblyAI-STT-3.7 ships with a new diarization model that has the lowest diarization error rate of the pre-recorded APIs AssemblyAI measured, 10.7% on English conversations. It costs $0.02/hr as an add-on.

How do I migrate from Deepgram to AssemblyAI for pre-recorded audio?

Replace model=nova-3 with speech_models: ["assemblyai-stt-3-7", "universal-2"], drop smart_format because formatting is on by default, and map diarize to speaker_labels. Set disfluencies to false if you want clean text. The AssemblyAI docs include a Deepgram migration guide that maps each parameter.

What languages does AssemblyAI support for batch transcription?

AssemblyAI-STT-3.7 covers 32 languages with native code switching. Adding Universal-2 to the speech_models list extends coverage to 99 languages, and Universal-2 is billed at $0.15/hr for that audio.