BroadRouter

Deepgram alternative for speech to text: 4 APIs compared

By the BroadNet Inference team · 7 min read · Published

Deepgram is a strong speech-to-text API with streaming, voice agents and a USD 200 free credit. People switch because requests join its Model Improvement Program by default, its API is not OpenAI-compatible, or for batch price and Arabic accuracy. AssemblyAI, OpenAI, Groq Whisper and BroadNet Transcribe are four alternatives with different trade-offs.

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

Deepgram is a mature speech platform: Nova-3 transcription, streaming, text to speech and a voice agent API, with a generous free credit. This page is for people with a specific reason to look elsewhere, whether that is data use, API format, price or Arabic. It compares four speech-to-text alternatives, including BroadNet Transcribe, and says when Deepgram is still the better choice.

What Deepgram does well

From Deepgram’s own pages, checked 2026-11-14.

  • Clear per-minute pricing. Nova-3 pre-recorded is USD 0.0043 a minute monolingual and USD 0.0052 multilingual on pay as you go; streaming is USD 0.0048 and USD 0.0058. The Growth plan, from USD 4,000 a year prepaid, lowers these by up to 20%.
  • Streaming and voice agents. Flux streaming STT, Aura-2 text to speech and a Voice Agent API at USD 0.075 a minute.
  • Languages. Nova-3 Multilingual covers 55 languages. Arabic is ar there, and monolingual Nova-3 lists regional codes such as ar-AE, ar-SA and ar-EG.
  • High concurrency. Pay as you go allows “up to 50 concurrent requests” for pre-recorded and “up to 150” for streaming.
  • A large free credit. USD 200 with no card, which does not expire.
  • Add-ons. Diarization is included with pre-recorded audio; on streaming, diarization, redaction and keyterm prompting are priced per minute.

Why people look for a Deepgram alternative

From Deepgram’s documentation, plus comparisons on the cited pricing pages.

  1. Training on your audio by default. “Requests participate in the Model Improvement Program (MIP) by default, and Deepgram retains them.” You exclude a request with mip_opt_out=true, and opted-out data is “retained only for the duration needed to process the request.” Every request has to carry the flag; one missed code path means retained audio. Request metadata and usage logs are retrievable for 90 days either way, and summarized usage is kept longer.
  2. Its own API format. Deepgram uses POST https://api.deepgram.com/v1/listen with Authorization: Token. Tools built for OpenAI’s speech-to-text endpoint need a separate integration.
  3. Price for batch audio. At USD 0.258 an hour for monolingual pre-recorded, Deepgram costs more than Groq Whisper (USD 0.04 to USD 0.111 an hour) and AssemblyAI Universal-2 (USD 0.15 an hour). The speech-to-text cost calculator compares them at your volume.
  4. Arabic dialects. Arabic is supported. If your audio is in Gulf, Egyptian or Levantine dialects, test a specialised Arabic model alongside Nova-3 before you commit. Measure with word error rate on your own clips.
  5. Growth needs a prepayment. The lower rates need at least USD 4,000 a year up front.

Deepgram alternatives compared

Deepgram AssemblyAI OpenAI Groq BroadNet Transcribe
Example price Nova-3 pre-recorded USD 0.0043/min (USD 0.258/hr) Universal-2 USD 0.15/hr gpt-transcribe USD 0.0045/min (USD 0.27/hr) Whisper large-v3-turbo USD 0.04/hr $0.002/min ($0.12/hr); Arabic $0.0042/min
Flat plan No; Growth is prepaid usage No No No No for speech (flat plans cover DeepSeek V4.1 Flash)
OpenAI-compatible No (/v1/listen) No for STT The reference Yes Yes
Data retention (cited) Retained for MIP by default; opt out per request Audio deleted within 24 to 48 hours; transcripts from 30 days Abuse logs up to 30 days Not retained by default; logs up to 30 days unless ZDR “keeps no prompts, messages, images, audio, responses or transcripts”
Languages 55 (Nova-3 Multilingual) 99 (Universal-2) Many Whisper’s languages 14, plus a dedicated Arabic model
Free credit USD 200, no card USD 50, no card Not listed Free plan with limits $5, no card

Sources for every cell are listed below. For more providers and add-on prices, see speech-to-text API pricing compared.

What 1,000 hours of audio costs

A worked example for batch (pre-recorded) transcription, at prices checked 2026-11-14 (OpenAI re-checked 2026-11-15). One hour is 60 minutes, so 1,000 hours is 60,000 minutes. OpenAI’s older gpt-4o-mini-transcribe (USD 0.18 an hour) is left out because OpenAI has deprecated it.

Provider and model Per hour 1,000 hours a month
Groq, Whisper large-v3-turbo USD 0.04 USD 40
Soniox, async (Arabic included) USD 0.10 USD 100
BroadNet Transcribe (14 languages) $0.12 $120.00
AssemblyAI, Universal-2 USD 0.15 USD 150
BroadNet Transcribe Arabic $0.25 $252.00
Deepgram, Nova-3 monolingual USD 0.258 USD 258
OpenAI, gpt-transcribe USD 0.27 USD 270
Deepgram, Nova-3 multilingual USD 0.312 USD 312

Price is only half the decision. A cheaper model that gets one word in five wrong on your audio costs more in human correction than it saves. Add-ons change the totals too: if you need diarization or redaction, check each provider’s add-on prices (Deepgram includes diarization on pre-recorded audio). And smaller files matter: Groq bills at least 10 seconds per request, which adds up on short voice notes. For 10,000 Arabic voice notes of 30 seconds each, BroadNet’s price works out to $21.00.

8 alternatives to Deepgram at a glance

Batch (pre-recorded) list prices, checked 2026-11-14. US-dollar prices run lowest first, then Gladia (priced in euros), then BroadNet’s own product.

  1. Groq (Whisper large-v3-turbo): USD 0.04 an hour, OpenAI-compatible, with a 10-second minimum per request. The lowest price we found.
  2. Soniox: USD 0.10 an hour for files and USD 0.12 for real-time, 60+ languages including Arabic. Cheaper per hour than both BroadNet Transcribe models.
  3. AssemblyAI (Universal-2): USD 0.15 an hour, 99 languages and a USD 50 free credit.
  4. ElevenLabs Scribe: USD 0.20 an hour on the API, 90+ languages; Arabic sits in its “Good” accuracy group.
  5. Speechmatics: USD 0.24 an hour list on Melia 1 (down to USD 0.12 with its model-training discount and top-tier annual commitment), 55+ languages with Arabic listed, and USD 100 of free credit with no card.
  6. OpenAI: gpt-transcribe is USD 0.0045 a minute (USD 0.27 an hour), checked 2026-11-15. The older gpt-4o-mini-transcribe (USD 0.003 a minute), gpt-4o-transcribe, gpt-4o-transcribe-diarize (speaker labels) and whisper-1 (USD 0.006) are marked deprecated, and OpenAI’s deprecations page says they shut down on February 26, 2027.
  7. Gladia: from EUR 0.16 an hour async on Solaria-3 (five languages) or EUR 0.27 on Solaria-1 (100+ languages), with EUR 50 of credit for new accounts, valid for a year.
  8. 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 Deepgram

  • You stream live audio for captions, call centres or voice bots. Streaming concurrency of 150 on pay as you go is high.
  • You build voice agents and want STT, TTS and orchestration from one vendor.
  • You need diarization, redaction or keyterm prompting. Diarization is included on pre-recorded audio.
  • You haven’t used the USD 200 credit yet. That covers tens of thousands of minutes at Nova-3 rates.

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.

  • Arabic is your main language. BroadNet Transcribe Arabic costs $0.0042 a minute ($0.25 an hour). 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; that is BroadNet’s report, so test it on your audio. See the Arabic speech-to-text guide.
  • You want no retention and no flag to remember. No audio or transcripts are kept on any request.
  • You use the OpenAI SDK or tools built for it. BroadNet Transcribe covers 14 languages on the same /v1/audio/transcriptions call.
  • You want an LLM on the same key to summarise or classify the transcript.

Limits to know: files are WAV or MP3 up to 25 MiB each, and BroadNet’s model pages describe file transcription only; they don’t list streaming, diarization or redaction. Groq Whisper and Soniox are cheaper per hour, and Soniox lists Arabic, so include it in your Arabic test.

When another alternative fits better

  • AssemblyAI for 99 languages, add-ons such as diarization (+USD 0.02 an hour) and PII redaction, and a USD 50 credit. Training on some files is possible “as permitted by the applicable contract”, with an opt-out.
  • Groq for the lowest-priced Whisper we found: USD 0.04 an hour on large-v3-turbo, OpenAI-compatible, with a 10-second minimum per request.
  • OpenAI if you already use its API: gpt-transcribe at USD 0.0045 a minute, its recommended transcription model. OpenAI’s only speaker-labelling model, gpt-4o-transcribe-diarize, is deprecated and shuts down on February 26, 2027.

How to switch from Deepgram

Deepgram’s /v1/listen call becomes an OpenAI-style transcription call:

from openai import OpenAI
import os

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

# Deepgram: POST https://api.deepgram.com/v1/listen?model=nova-3&language=ar
with open("voice-note.mp3", "rb") as f:
    result = client.audio.transcriptions.create(
        model="broadnet/cohere-transcribe-arabic",  # or broadnet/cohere-transcribe for 14 languages
        file=f,
    )
print(result.text)
  1. Pick 30 to 50 real clips per language, with human transcripts.
  2. Run them through Deepgram (with mip_opt_out=true) and each alternative. Compare word error rate; Arabic WER explained shows how to normalise Arabic text first.
  3. Convert other formats to WAV or MP3 and split files over 25 MiB. Split on silence rather than at fixed offsets, so no word is cut in half at a boundary.
  4. Move batch transcription first; keep Deepgram for streaming if you need it.

The short version

Deepgram is a good choice for streaming and voice agents, and its free credit is large. If default training on your audio, a non-OpenAI API or Arabic accuracy is the issue, test alternatives on your own clips. BroadNet Transcribe fits Arabic-heavy, privacy-sensitive batch work; Groq fits lowest-cost Whisper. For WhatsApp audio, see transcribing voice notes.

Questions

What is the best Deepgram alternative?

For broad language coverage and add-ons, AssemblyAI (Universal-2 covers 99 languages at USD 0.15 an hour). For the lowest price on Whisper, Groq. For drop-in OpenAI SDK compatibility, zero retention and a dedicated Arabic model, BroadNet Transcribe. For real-time voice agents, Deepgram is still hard to replace.

Does Deepgram use my audio to train models?

By default, requests participate in its Model Improvement Program and are retained. Add mip_opt_out=true to each request to exclude it; opted-out audio and transcripts are kept only as long as needed to process the request. Request metadata and usage logs are retrievable for 90 days, and summarized usage is kept longer.

Is Deepgram compatible with the OpenAI transcription API?

No. Deepgram uses POST https://api.deepgram.com/v1/listen with a Token authorization header. OpenAI, Groq and BroadNet Inference use the OpenAI /v1/audio/transcriptions format, so the OpenAI SDK works with them by changing the base URL.

Which speech-to-text API is best for Arabic?

Test on your own audio. Deepgram Nova-3 lists Arabic codes, AssemblyAI puts Arabic in its "good accuracy" group, and BroadNet reports its Arabic model 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. Measure word error rate on 30 to 50 of your own clips.

How much free credit does Deepgram give?

Deepgram's pricing page offers a USD 200 credit with no card and no expiry. AssemblyAI offers USD 50 and BroadNet Inference $5, which covers about 1,190 minutes of Arabic transcription.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity