The OpenAI Whisper API is the default many teams start with: one endpoint, one price, dozens of languages. You look for a Whisper API alternative when one of four things happens: accuracy on your audio is not good enough, the bill grows with volume, your security team wants less data kept, or the model you call is being retired. This post shows how to decide, and how to test an alternative without rewriting your code.
Is the OpenAI Whisper API being retired?
Yes, on OpenAI’s current schedule. Its deprecations page, checked 2026-11-15, lists whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize as deprecated, with notice given on August 26, 2026 and shutdown on February 26, 2027. The recommended replacement is gpt-transcribe (USD 0.0045 a minute, or USD 0.27 an hour) for files, or gpt-live-transcribe for live audio. Its pricing page marks the same four models deprecated, and its speech to text guide says to “Start with gpt-transcribe”. The /v1/audio/translations endpoint is deprecated too. So you will re-test your audio on a new model either way. That makes now a good time to compare OpenAI’s replacement with other providers on the same clips.
For the full picture, our open speech-to-text API price index lists batch and streaming prices, billing increments and Arabic support for more than 20 products, each checked against the provider’s own pricing page (CSV and JSON).
Why do teams look for a Whisper API alternative?
| Reason | What to check in an alternative |
|---|---|
| Accuracy on a specific language or dialect | Published word error rate on a relevant test set, then your own test |
| Price at volume | Price per audio hour and billing minimums |
| Data retention | Whether audio and transcripts are stored, and for how long |
| Limits | File size, formats, rate limits |
| Features | Timestamps, diarization, prompts, streaming |
OpenAI’s speech to text guide says “Whisper supports 98 languages, but accuracy varies by language”. That is the honest starting point: the language list tells you what is possible, not what is good.
Which Whisper API alternatives are there, and what do they cost?
| Option | Price (list, Nov 2026) | Languages | Notes |
|---|---|---|---|
OpenAI Whisper API (whisper-1) |
USD 0.006/min | 98 (accuracy varies) | 25 MB files; prompts up to 224 tokens; shuts down 2027-02-26 |
OpenAI gpt-transcribe |
USD 0.0045/min (USD 0.27/hour) | See OpenAI’s guide | OpenAI’s recommended replacement; OpenAI’s guide shows word timestamps only for whisper-1 and speaker labels only for the deprecated diarize model |
| Groq, Whisper Large v3 | USD 0.111/hour | Multilingual | Groq states 10.3% WER; 10-second billing minimum |
| Groq, Whisper Large v3 Turbo | USD 0.04/hour | Multilingual | Groq states 12% WER; not trained for translation |
| Soniox (async) | about USD 0.10/hour, billed per token | More than 60, Arabic listed | Own job-based API, not OpenAI-compatible; real-time about USD 0.12/hour |
| Self-hosted Whisper | Your GPU cost | Same as Whisper | MIT licence; you run and scale it |
| BroadNet Transcribe | $0.002/min | 14 | Not Whisper; based on Cohere Transcribe |
| BroadNet Transcribe Arabic | $0.0042/min | Arabic | BroadNet reports #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 |
Prices and Groq’s figures were checked on official pages on 2026-11-14; OpenAI’s were re-checked on 2026-11-15. OpenAI’s gpt-transcribe costs more per hour than both BroadNet models. Groq’s Turbo endpoint and Soniox both cost less per hour than BroadNet Transcribe ($0.12 an hour); Soniox needs new client code because it does not use the OpenAI request shape. Groq’s WER figures are its own published numbers, not a test on a common set with the others, so do not compare them directly with BroadNet’s. This page is about switching; for the full price list, including DeepSeek, AssemblyAI, Deepgram, Google, AWS and Speechmatics with their add-on prices, see the speech-to-text pricing comparison, and the speech-to-text cost calculator prices your own monthly hours.
What do the accuracy numbers really mean?
BroadNet publishes three WER figures for BroadNet Transcribe Arabic in its model catalogue:
| Test set | Kind of audio | WER |
|---|---|---|
| FLEURS-ar | Clean read sentences | 3.8 |
| MASC-ar | Mixed real-world Arabic | 21.1 |
| BroadNet dialect set | Dialect speech | 23.4 |
On the same MASC and dialect sets, BroadNet reports Whisper large-v3 at 31.5 and 31.8, against 21.1 and 23.4 for its Arabic model: about a third fewer errors on MASC and about a quarter fewer on the dialect set. That is BroadNet’s own test of a competitor, so check it on your audio. On the public Open Universal Arabic ASR Leaderboard, which uses the same test sets for every entry, Whisper large-v3 was #19 of 39 on 15 November 2026 (average WER 36.86), and Cohere Transcribe Arabic 07-2026, the model BroadNet serves, was #4 (24.19).
The spread is the lesson. The same model is near-perfect on read speech and much weaker on conversation. Any provider’s headline WER tells you how it did on someone else’s audio. Dialects are covered in depth in Arabic dialect speech recognition.
How to test a Whisper alternative in one afternoon
- Collect 30 to 50 clips per language, from real traffic: calls, voice notes, meetings. Remove anything you may not share.
- Write reference transcripts for 10 clips per language by hand.
- Run both APIs on the same clips. With OpenAI-compatible providers, only the client and model change:
import os
from openai import OpenAI
whisper = OpenAI() # uses OPENAI_API_KEY
broadnet = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])
def run(client, model, path, lang):
with open(path, "rb") as f:
return client.audio.transcriptions.create(model=model, file=f, language=lang).text
hyp_a = run(whisper, "whisper-1", "clip01.wav", "ar") # also try "gpt-transcribe", OpenAI's replacement
hyp_b = run(broadnet, "broadnet/cohere-transcribe-arabic", "clip01.wav", "ar")
- Score with the same normalisation for both, using jiwer:
import re, jiwer
def norm(t):
t = re.sub(r"[ً-ْ]", "", t) # strip Arabic diacritics
t = re.sub(r"[إأآ]", "ا", t).replace("ى", "ي") # unify alef and yaa
return re.sub(r"[^\w\s]", "", t).lower().strip()
print(jiwer.wer(norm(reference), norm(hyp_a)), jiwer.wer(norm(reference), norm(hyp_b)))
- Read 20 transcripts yourself. Check numbers, names and product terms. A model can win on WER and still get the account number wrong.
When should you self-host Whisper instead?
Whisper’s code and weights are MIT-licensed, so running it yourself is always an option. It makes sense when:
- Audio may not leave your network, for legal or contractual reasons.
- Your volume keeps a GPU busy most of the day, so you are not paying for idle hours.
- You need to fine-tune on your own vocabulary or accent.
It costs more than it looks when traffic is spiky, when you need high availability across several machines, or when nobody on the team wants to own GPU drivers and upgrades (the GPU break-even math). The openai/whisper README also notes that the turbo model is not trained for translation. A hosted API turns those problems into a per-minute price, which is why most teams start there and only self-host once volume and requirements are clear.
Migration checklist
Changing the base URL is the easy part. Before you switch traffic:
- Formats. OpenAI accepts flac, mp3, mp4, m4a, ogg, wav and webm. BroadNet accepts WAV and MP3, so convert
.oggand.m4awith ffmpeg first. For browser recordings, see record audio in the browser and send it to a speech-to-text API. - File size. OpenAI caps uploads at 25 MB; BroadNet at 26,214,400 bytes (25 MiB).
- Language field. Always send it. On BroadNet, Arabic without
language=arcan come back in Latin letters. - Optional parameters. If you rely on
prompt,verbose_jsonor word timestamps, confirm the alternative supports them. BroadNet’s docs describe a JSON response withtext. - Rate limits. BroadNet pay-as-you-go keys allow 30 requests a minute and 4 at once.
The general version of this list is in the OpenAI-compatible migration checklist.
Where BroadNet Transcribe fits
Choose BroadNet when Arabic accuracy matters, when you want audio and transcripts not retained (zero data retention), or when $0.002 a minute across 14 languages beats your current bill. The $5 starting credit covers about 1,190 minutes of Arabic audio, enough to run the test above; get a key at BroadNet Inference. The Arabic speech-to-text guide has curl and Python examples.
Stay with OpenAI’s replacement models, or another provider, if you need one of the many languages BroadNet does not list, speech translation, live transcription, or prompts and keyword hints (OpenAI’s guide documents them for gpt-transcribe). For translating Arabic recordings to English from files, a two-step pipeline also works: see translate Arabic audio to English. For the very lowest English price, hosted Whisper Turbo endpoints cost less per hour.
