BroadRouter

BroadNet Transcribe vs BroadNet Transcribe Arabic

Updated

BroadNet Transcribe costs $0.002 per audio minute and BroadNet Transcribe Arabic $0.0042. Pick BroadNet Transcribe for speech in any of 14 languages at BroadNet’s lowest per-minute price; pick BroadNet Transcribe Arabic for Arabic audio: voice notes, calls and dialects. Both work on the free starting credit, with zero data retention.

Both models answer on the same OpenAI-compatible /v1/audio/transcriptions endpoint, accept WAV or MP3 up to 26,214,400 bytes per file and return plain text. BroadNet lists both as based on Cohere Transcribe. The general model covers 14 languages at the lower per-minute price; the Arabic model is the one to use for Arabic speech, including dialects. Since both return text only, the decision is language and price, not features.

Side by side

BroadNet Transcribe compared with BroadNet Transcribe Arabic
BroadNet TranscribeBroadNet Transcribe Arabic
Model IDbroadnet/cohere-transcribebroadnet/cohere-transcribe-arabic
Based onCohere TranscribeCohere Transcribe
Price per audio minute$0.002$0.0042
Price per hour of audio$0.12$0.25
Languages14 languages, Arabic includedArabic
Audio formatsWAV or MP3, up to 25 MiB a fileWAV or MP3, up to 25 MiB a file
OutputPlain text: {"text": ...}Plain text: {"text": ...}
Endpoint/v1/audio/transcriptions/v1/audio/transcriptions
Works on free creditYesYes

Prices in USD from BroadNet's rate card, synced 2026-11-14. "Not listed" means BroadNet's docs do not list the feature for that model, not that it is absent.

Report a wrong price or number (email to info@broadnet.ai, with this page's link filled in).

What the same audio costs

Both models bill per minute of audio, whatever the language or the length of the transcript. Each row prices one workload on both models; the multiplier column shows how many times more the dearer model costs. Try other volumes in the speech-to-text cost calculator.

WorkloadHoursTranscribeTranscribe ArabicMultiplier
Voice notes: 10,000 messages of 30 seconds83$10.00$21.002.1×
Podcast or lecture archive: 100 hours100$12.00$25.202.1×
Call centre: 20 agents, 6 hours of calls a day, 22 days2,640$316.80$665.282.1×
Recording backlog: 10,000 hours10,000$1,200.00$2,520.002.1×

The $5 starting credit covers 2,500 minutes on Transcribe or 1,190 minutes on Transcribe Arabic, enough to test a real batch of recordings before paying.

From audio to tokens: the next step's cost

A transcript is usually the input to a chat model that summarises, tags or answers questions about it. Assume a speaker says 120 words a minute (an assumption; measure your own audio), so an hour of speech is 7,200 words. Using our measured tokens per word, this is what one hour of transcript costs to send to each chat model on BroadNet, next to the cost of transcribing it.

Chat modelArabic tokens / hourInput costEnglish tokens / hourInput cost
DeepSeek V4.1 Flash15,989$0.0039,172$0.002
GLM 5.317,564$0.0149,288$0.007
Qwen3.8 27B13,173$0.0029,501$0.002
BroadNet Gemma 4 VL 26B14,593$0.0019,355$0.001

Transcribing that hour costs $0.12 on Transcribe and $0.25 on Transcribe Arabic. So the transcription, not the summary, is most of the bill for a voice-note or call pipeline. Count real transcripts with the Arabic token counter.

Context, file limits, plans and keys

EncodingBytes per secondMinutes per file
WAV, 16 kHz, 16-bit, mono (what BroadNet’s ffmpeg example produces)32,00013.7
WAV, 44.1 kHz, 16-bit, stereo (CD quality)176,4002.5
MP3 at 128 kbit/s16,00027.3
MP3 at 64 kbit/s8,00054.6
MP3 at 32 kbit/s (speech)4,000109.2

The per-file limit is 26,214,400 bytes, so the duration that fits depends on the encoding. BroadNet's docs convert phone and WhatsApp audio with ffmpeg -i in.ogg -ar 16000 -ac 1 out.wav, which produces the first row; Cohere's cards say the model resamples to 16 kHz mono anyway, so a higher sample rate only makes the file bigger. Both models use the same pay-as-you-go key (30 requests a minute, 4 in flight), are not part of the unlimited plans, and process audio with zero data retention. See the WhatsApp voice notes guide for the full pipeline.

For recordings longer than one file allows, cut them into fixed segments before upload. At 16 kHz mono WAV (32,000 bytes a second), a 10-minute segment is 19,200,000 bytes, safely under the limit, and a 10,000-hour backlog becomes 60,000 requests. At the pay-as-you-go limit of 30 requests a minute, sending them takes at least 34 hours, and longer if requests take more time than the 4-in-flight limit allows.

ffmpeg -i call.mp3 -f segment -segment_time 600 -ar 16000 -ac 1 part%03d.wav

Cohere's cards note that the models are "eager to transcribe, even non-speech sounds," and recommend a noise gate or voice activity detection step before transcription, so trimming long silences also saves billed minutes. Read word error rate to compare accuracy figures correctly.

Worked example: A bilingual call archive in the Gulf

A contact centre in the Gulf records 5,000 hours of calls a month. About 60% are in Arabic and 40% in English. It can send everything to one model, or detect the language per call (for example from the dialled line or the agent’s queue) and route each call to the matching model.

Option for 5,000 hours a monthCost per month
Everything on Transcribe$600.00
Everything on Transcribe Arabic$1,260.00
Arabic to Transcribe Arabic, the rest to Transcribe$996.00

Routing sits between the two options: it costs more than sending everything to the general model and less than sending everything to the Arabic model, and every Arabic call gets the Arabic model. Because Cohere’s cards say the models do not detect language and perform inconsistently on code-switched speech, routing by a known language is also the more accurate setup.

Strengths and weaknesses, with sources

Each point cites a public model card or BroadNet's own documentation, checked 2026-11-14. Points about token counts come from our measurement above.

Where BroadNet Transcribe is stronger

  • Covers 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese and Arabic. (BroadNet docs, Cohere card)
  • The lower per-minute price on BroadNet, less than half of the Arabic model’s rate. (BroadNet docs)
  • Cohere describes the base model as a 2B-parameter Conformer encoder-decoder, trained from scratch for transcription accuracy and speed. (Cohere card)

Where BroadNet Transcribe Arabic is stronger

  • BroadNet reported it #2 of 37 on the Open Universal Arabic ASR Leaderboard in September 2026, and #1 among models whose licences permit API serving. When we checked the live board on 15 November 2026, it listed the model #4 of 39. (BroadNet Transcribe Arabic page, leaderboard)
  • On BroadNet’s own benchmark it reports 21.1 word error rate on MASC dialect speech, against 22.3 for ElevenLabs Scribe v2 and 31.5 for Whisper large-v3. (BroadNet Transcribe Arabic page)
  • BroadNet’s lab measurements report word error rates of 3.8 on FLEURS Arabic, 21.1 on MASC Arabic and 23.4 on dialect speech. (BroadNet models)
  • BroadNet’s model page says it serves Cohere Transcribe Arabic 07-2026 (2B parameters, Apache-2.0), which Cohere’s card says is optimized for Arabic, Arabic dialects, English and Arabic–English code-switched speech. (BroadNet Transcribe Arabic page, Cohere Arabic card)

Published benchmark numbers

BenchmarkTranscribeTranscribe Arabic
Open Universal Arabic ASR Leaderboard rank#8 of 39 (live board, 15 November 2026) [leaderboard]#4 of 39 (live board, 15 November 2026); BroadNet reported #2 of 37 in September 2026 [leaderboard]
FLEURS Arabic, word error ratenot published3.8 [BroadNet Transcribe Arabic page]
MASC Arabic, word error ratenot published21.1 [BroadNet Transcribe Arabic page]
Arabic dialect speech, word error ratenot published23.4 [BroadNet Transcribe Arabic page]
FLEURS English, word error rate5.5 [BroadNet Transcribe page]not published
FLEURS Spanish, word error rate3.2 [BroadNet Transcribe page]not published
Open ASR Leaderboard (English), mean word error rate5.42 (Cohere base model) [Cohere card]not published

Word error rate is lower-is-better. The Arabic figures are BroadNet’s own lab measurements and BroadNet’s reported leaderboard rank for its Arabic model. BroadNet also publishes FLEURS English and Spanish figures for BroadNet Transcribe (Cohere Transcribe 03-2026). The Open ASR Leaderboard figure is from Cohere’s card for the same base model. The Arabic leaderboard ranks both Cohere models, on the same test sets: the Arabic model well ahead of the general one. Otherwise no source measures both models on the same test set. Figures checked 2026-11-14. Before switching, evaluate both on your own data.

Which should you choose?

Pick BroadNet Transcribe if…

  • your audio is in one of the 13 non-Arabic languages, or mostly English.
  • price per hour is the deciding factor and the audio is clean, single-language speech.
  • you transcribe a multilingual archive and want one model for every language.

Pick BroadNet Transcribe Arabic if…

  • the audio is Arabic, including dialect speech.
  • you transcribe WhatsApp voice notes or call-centre recordings from Arabic-speaking customers.
  • better Arabic accuracy is worth about twice the general model’s per-minute price.

Or use both

If you know the language of each file, route Arabic to BroadNet Transcribe Arabic and everything else to BroadNet Transcribe, and always send the language field. Cohere’s cards say the models do not detect language automatically and perform inconsistently on code-switched audio, so do not leave language out.

Before you switch

  • Only WAV and MP3 are accepted. Convert WhatsApp .ogg and phone .m4a files with ffmpeg first.
  • Billing is per audio minute, silence included. Trimming long silences, or running voice activity detection first, lowers the bill and, per Cohere’s cards, reduces made-up text from background noise.
  • Send language on every request, including on the 14-language model. Without it, Arabic can come back transliterated into Latin letters.

Switch with one line

Both models answer on https://inference.broadnet.ai/v1 with the same key. Change only the model field:

client.audio.transcriptions.create(model="broadnet/cohere-transcribe", file=f, language="ar")  # or "broadnet/cohere-transcribe-arabic"

In short: Transcribe is best for speech in any of 14 languages at BroadNet’s lowest per-minute price; Transcribe Arabic for Arabic audio: voice notes, calls and dialects. New to the API? Start with the Python SDK guide or Node guide, or compare every model.

Try both on $5 of free creditSee all prices

Questions

Is BroadNet Transcribe cheaper than BroadNet Transcribe Arabic?

Yes. BroadNet Transcribe costs $0.002 per audio minute against $0.0042, so an hour costs $0.12 against $0.25, and 1,000 hours cost $120.00 against $252.00.

Which one should I use for Arabic?

For Arabic audio, BroadNet Transcribe Arabic. It is the Arabic-specific model, reported by BroadNet at #2 of 37 on the Open Universal Arabic ASR Leaderboard in September 2026 (the live board listed it #4 of 39 on 15 November 2026). Always send language=ar; without it, Arabic can come back transliterated into Latin letters.

How long can one audio file be?

The limit is 26,214,400 bytes per file, not a duration. That is about 13 minutes of 16 kHz mono WAV, 27 minutes of 128 kbit/s MP3 or 109 minutes of 32 kbit/s MP3. Split longer recordings into chunks.

How many minutes does the $5 free credit cover?

2,500 minutes on BroadNet Transcribe or 1,190 minutes on BroadNet Transcribe Arabic, about 41.7 and 19.8 hours. Both work on the free credit without a top-up.

Do BroadNet Transcribe models detect the language automatically?

Do not rely on it. Cohere’s cards say the models have no explicit automatic language detection and work best on a single, pre-specified language, and BroadNet’s docs say to always send language: language=ar for Arabic, language=en for English. Without it, Arabic can come back transliterated into Latin letters.

Do these models return timestamps or speaker labels?

No. BroadNet’s endpoint returns the transcript as {"text": ...}, and Cohere’s model cards state that Cohere Transcribe has no timestamps or speaker diarization. Use a separate diarization step if you need speakers.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity