BroadRouter

AssemblyAI alternative: pricing, retention and Arabic compared

By the BroadNet Inference team · 8 min read · Published

AssemblyAI is a full speech AI platform: 99 languages, streaming, diarization, PII redaction and an LLM Gateway, from USD 0.15 an audio hour plus add-ons. People look elsewhere for an OpenAI-style transcription call, cheaper Whisper, no model training on free-tier audio, or zero retention. Deepgram, OpenAI, Groq, Soniox and BroadNet Transcribe are the main alternatives.

BroadRouter is run by BroadNet, which sells BroadNet Inference, one of the options compared here. Facts about AssemblyAI and other providers are quoted from their official pages, with links in the sources. About BroadRouter · editorial policy.

AssemblyAI is one of the most complete speech AI APIs: batch and streaming transcription, speaker labels, redaction, summaries and an LLM Gateway, all on one key. This page is for people who have a reason to look elsewhere, whether that is the API shape, how add-ons add up, how data is kept, or Arabic. It compares the main alternatives, including BroadNet’s own, and says plainly when AssemblyAI is still the better pick. For a Deepgram-focused view, see the Deepgram alternative page.

What does AssemblyAI do well?

From AssemblyAI’s own pages, checked 2026-11-15.

  • Low base prices. Universal-2 is USD 0.15 an hour for pre-recorded audio and Universal-3.5 Pro is USD 0.21, both billed “to the exact second”. Failed transcripts are not charged.
  • Languages. Universal-2 covers 99 languages. Universal-3.5 Pro covers 18, and can be set to fall back to Universal-2 for the rest. A Universal-3.6 Pro Preview covers 32.
  • A deep feature list. Diarization (+USD 0.02 an hour on pre-recorded), PII audio redaction (+USD 0.05), PII text redaction (+USD 0.08), sentiment, entity detection, translation, summaries and medical mode.
  • Streaming. Universal-Streaming is USD 0.15 an hour; Universal-3.6 Pro Realtime is USD 0.45 an hour for voice agents.
  • Large files and URLs. Its migration guide says it “supports file sizes up to 5GB” and can transcribe a public audio URL directly.
  • High parallel limits. Paid accounts get “200+” transcriptions processing at once; free accounts get 5.
  • A free credit. USD 50 for new accounts, with no expiry. It does not cover the LLM Gateway.

Why do people look for an AssemblyAI alternative?

Each point is from AssemblyAI’s documentation.

  1. Its own API format. You authenticate with your key in an authorization header “(no Bearer prefix)”, submit to POST /v2/transcript, then poll GET /v2/transcript/{id} until the status is completed, or use a webhook. Polling is capped at 60 requests a minute per transcript. Its Sync API (POST /v1/transcribe, up to 20 minutes) skips polling but is not OpenAI’s format either. Code written for OpenAI’s speech-to-text endpoint needs a separate integration. AssemblyAI’s own OpenAI-compatible endpoints are for its LLM Gateway and text to speech; its transcription API uses its own format.
  2. Add-ons stack. Each feature is a separate hourly rate on top of the model. A call-centre hour on Universal-2 with diarization and PII text redaction is USD 0.15 + 0.02 + 0.08 = USD 0.25. Entity detection (+USD 0.08) or topic detection (+USD 0.15) push it higher.
  3. Model training. “Only certain files submitted to the API, as permitted by the applicable contract, are used for model training”, after a PII redaction step. Paid users can opt out from the Data Controls page, and opt-outs apply only to new requests. “Free users do not have the ability to opt out.”
  4. Retention on batch audio. Without a TTL or a BAA, uploaded audio is deleted “at 24 hours and is at most 48 hours”, and transcripts from 30 days. You can set a TTL as low as one hour or delete transcripts through the API. Zero data retention is offered for streaming, Sync STT and Dictation when you are opted out of training.
  5. Arabic accuracy. Universal-2 lists Arabic under “Good accuracy (>10% to ≤25% WER)”, not its high-accuracy group. Universal-3.5 Pro includes Arabic, but AssemblyAI’s dialect notes cover English, Spanish, French, Portuguese and Italian only. Measure word error rate on your own audio.

AssemblyAI alternatives compared

AssemblyAI Deepgram OpenAI Groq BroadNet Transcribe
Batch price Universal-2 USD 0.15/hr; Universal-3.5 Pro USD 0.21/hr Nova-3 USD 0.0043/min (USD 0.258/hr) gpt-transcribe USD 0.0045/min; gpt-4o-mini-transcribe USD 0.003/min (deprecated) Whisper large-v3-turbo USD 0.04/hr $0.002/min ($0.12/hr); Arabic $0.25/hr
Diarization +USD 0.02/hr Included on pre-recorded gpt-4o-transcribe-diarize, USD 0.006/min (deprecated, shuts down 2027-02-26) Not documented Not listed
Streaming Yes Yes Yes Not documented Not listed
OpenAI-compatible transcription No (/v2/transcript) No (/v1/listen) The reference Yes Yes
Max file 5 GB, or a URL See its docs See its docs 25 MB free, 100 MB dev; URL for larger WAV or MP3, 25 MiB
Training and retention Some files may train; opt-out for paid users; audio deleted in 24 to 48 hrs May use requests in its Model Improvement Program; opt out per request (mip_opt_out=true) See its data controls Not retained by default “keeps no prompts, messages, images, audio, responses or transcripts”
Free credit USD 50, no expiry USD 200, no card Not listed Free tier $5, no card

Sources for each cell are listed below. To price your own volume, use the speech-to-text cost calculator.

AssemblyAI vs Deepgram

These are the two names most people shortlist. On base price for batch audio, AssemblyAI is cheaper: USD 0.15 an hour on Universal-2 against USD 0.258 for Deepgram Nova-3 monolingual and USD 0.312 multilingual. Deepgram includes diarization on pre-recorded audio, charges USD 0.0020 a minute (USD 0.12 an hour) for redaction, and gives a USD 200 credit to AssemblyAI’s USD 50. Both use their own transcription API. On training, Deepgram may use requests in its Model Improvement Program unless you opt out per request; AssemblyAI uses an account-level opt-out that free users don’t get. Both list Arabic.

The LLM layer: summarising transcripts

AssemblyAI’s LLM Gateway lists 60 models behind an OpenAI-style /v1/chat/completions endpoint, and AssemblyAI says it “has opted out of data training with all LLM Gateway providers”. In-region US or EU endpoints cost 10% more than global routing. DeepSeek V4.1 Flash lists at USD 0.30 input and USD 1.20 output per 1M tokens there. BroadNet Inference serves the same model at $0.175 and $0.52, on the same key as its transcription models. Compare per-token costs with the LLM API cost calculator.

7 alternatives to AssemblyAI at a glance

Batch list prices, checked 2026-11-15, lowest first, then BroadNet’s own product.

  1. Groq (Whisper large-v3-turbo): USD 0.04 an hour, OpenAI-compatible, billed at least 10 seconds per request. No diarization documented.
  2. Soniox: USD 0.10 an hour for files and USD 0.12 for real-time, with diarization and language identification “bundled into the hourly rate”. It lists 60+ languages, Arabic included.
  3. ElevenLabs Scribe: USD 0.20 an hour for Scribe v3 and v2, 90+ languages, diarization up to 32 speakers. Arabic sits in its “Good” accuracy group.
  4. Speechmatics: USD 0.24 an hour list on Melia 1, a 33% discount if you opt into its model-training programme, and USD 100 of free credit with no card.
  5. Deepgram: Nova-3 at USD 0.258 an hour monolingual, diarization included on pre-recorded audio, Arabic codes such as ar-AE, ar-SA and ar-EG, and a USD 200 credit.
  6. OpenAI: gpt-transcribe at USD 0.0045 a minute (USD 0.27 an hour). The older gpt-4o-mini-transcribe (USD 0.003), gpt-4o-transcribe-diarize (USD 0.006, speaker labels), gpt-4o-transcribe and whisper-1 are marked deprecated; OpenAI’s deprecations page says they shut down on February 26, 2027.
  7. BroadNet Transcribe: $0.12 an hour for 14 languages and $0.25 for the dedicated Arabic model, OpenAI-compatible, with no audio or transcripts kept. This is BroadNet’s own product.

When to stay with AssemblyAI

  • You need diarization, redaction or summaries in one request. BroadNet’s model pages list none of these.
  • You stream live audio for captions or voice agents.
  • Your files are long. 5 GB per file and direct URL input beat a 25 MiB limit.
  • You run big batches. 200+ parallel jobs on a paid account is far above BroadNet’s 4 requests in flight on pay as you go.
  • Arabic for less per hour. Universal-2 (USD 0.15) and Universal-3.5 Pro (USD 0.21) both cost less per hour than BroadNet Transcribe Arabic.

When BroadNet Transcribe fits

BroadNet Transcribe runs on BroadNet Inference, which hosts on BroadNet-owned or leased hardware in the United States, with partner capacity as backup.

  • You want no retention and no training, on every account. BroadNet’s privacy statement says it “keeps no prompts, messages, images, audio, responses or transcripts”, with no toggle to remember. See zero data retention.
  • Arabic is your main language. BroadNet Transcribe Arabic covers Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic, per BroadNet. BroadNet reports it ranked #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard. Rankings move, so test it on your own clips. The Arabic speech-to-text guide shows how.
  • You use the OpenAI SDK. BroadNet Transcribe covers 14 languages on the standard /v1/audio/transcriptions call, with no polling.
  • You want an LLM on the same key to summarise or tag the transcript.

Limits to know: WAV or MP3 only, up to 25 MiB per file; file transcription only, with no streaming, diarization or redaction listed. Groq and Soniox are cheaper per hour than both BroadNet models, and AssemblyAI is cheaper than BroadNet for Arabic.

When another alternative fits better

  • Groq for low-cost Whisper on short English or common-language clips. Our Whisper API alternative post covers it in depth.
  • Soniox if you want diarization included in a low hourly rate.
  • Deepgram for real-time voice agents and the larger free credit.
  • OpenAI if you already use its API and SDK: gpt-transcribe at USD 0.0045 a minute. OpenAI’s guide documents speaker labels only on the deprecated gpt-4o-transcribe-diarize.

How to switch from AssemblyAI

AssemblyAI’s upload, submit and poll steps become one OpenAI-style call:

from openai import OpenAI
import os

client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])

# AssemblyAI: POST /v2/upload, POST /v2/transcript, then poll GET /v2/transcript/{id}
with open("call.mp3", "rb") as f:
    result = client.audio.transcriptions.create(
        model="broadnet/cohere-transcribe-arabic",  # or broadnet/cohere-transcribe for 14 languages
        file=f,
        language="ar",
    )

summary = client.chat.completions.create(
    model="broadnet/deepseek-v4.1-flash",
    messages=[{"role": "user", "content": "Summarise this call in three bullets:\n" + result.text}],
)
print(summary.choices[0].message.content)
  1. List the AssemblyAI features you actually use. If you rely on diarization or redaction, keep AssemblyAI for those jobs or pick Soniox or Deepgram.
  2. Pick 30 to 50 real clips per language with human transcripts and compare WER. Arabic WER explained shows how to normalise Arabic first.
  3. Convert other formats to WAV or MP3 and split files over 25 MiB on silence.
  4. Move plain batch transcription first, and keep streaming where it is.

The short version

AssemblyAI is strong on features, languages and limits, and its base price is low; add-ons are where the bill grows. If you want an OpenAI-compatible transcription call, zero retention without an opt-out, or a dedicated Arabic model, test BroadNet Transcribe. If price per hour is all that matters, Groq and Soniox cost less. For call recordings, see transcribing call centre recordings.

Questions

What is the best AssemblyAI alternative?

It depends on why you are leaving. Deepgram for real-time voice agents and a USD 200 credit. Groq for low-cost Whisper (USD 0.04 an hour). Soniox for low-cost files with diarization included. BroadNet Transcribe for an OpenAI-compatible call with zero retention and a dedicated Arabic model.

How much does AssemblyAI cost?

Pre-recorded audio is USD 0.15 an hour on Universal-2 and USD 0.21 on Universal-3.5 Pro, billed to the second. Streaming starts at USD 0.15 an hour. Add-ons are extra per hour, for example diarization USD 0.02, PII text redaction USD 0.08 and entity detection USD 0.08. New accounts get USD 50 of credit that does not expire.

AssemblyAI vs Deepgram, which is cheaper?

For batch audio, AssemblyAI. Universal-2 is USD 0.15 an hour; Deepgram Nova-3 pre-recorded is USD 0.0043 a minute (USD 0.258 an hour) monolingual. Deepgram includes diarization on pre-recorded audio and gives a larger USD 200 credit, so compare with the add-ons you actually need.

Does AssemblyAI use my audio to train models?

It may. Its docs say only certain files, "as permitted by the applicable contract", are used for training, after PII redaction. Files are excluded under a BAA, on its European servers, or if you opt out. Paid users opt out in the dashboard; free users cannot opt out.

Does AssemblyAI support Arabic?

Yes. Arabic (ar) is one of the 18 languages of Universal-3.5 Pro, and Universal-2 lists it in its "Good accuracy" group (over 10% to 25% WER). Test it against a dedicated Arabic model such as BroadNet Transcribe Arabic on your own clips.

Is AssemblyAI compatible with the OpenAI API?

Partly. Its LLM Gateway uses POST /v1/chat/completions and works with the OpenAI SDK. Its transcription API does not: you upload or pass a URL to /v2/transcript, then poll. OpenAI, Groq and BroadNet accept the OpenAI /v1/audio/transcriptions call.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity