To transcribe calls every week, use the ElevenLabs Scribe v2 API.

Why ElevenLabs Scribe v2

The reasons it wins for this job.

Its 2.2% word error rate on Artificial Analysis's benchmark is among the best available through a mature public API.
Diarization (diarize=true) labels speakers, so the downstream summary knows which words were the prospect's.
It is priced per hour of audio in the low cost tier, and third-party reporting says ElevenLabs cut speech-to-text prices by up to 45% at the end of August 2026.
A realtime version exists on the same platform if you later want live notes during the call.

How to use it

4 steps to a first result.

1Pull recordings from your call toolMost meeting and dialler tools expose recordings through a webhook or API.
2Transcribe with diarizationSend each file to Scribe v2 with speaker labels on.
3Summarise into CRM fieldsSend the transcript to a text model with a JSON schema: pain, objections, next step, owner.
4Write to the CRM as a draft noteLet the rep confirm before fields overwrite anything.

ElevenLabs API, curl

Start from this.

Edit the parts in capitals, then run it.

curl -X POST https://api.elevenlabs.io/v1/speech-to-text \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -F model_id=scribe_v2 \
  -F diarize=true \
  -F file=@call-recording.mp3

Nothing is sent anywhere. It copies to your clipboard.

Alternatives that also work

If ElevenLabs Scribe v2 is not an option.

Deepgram Nova-3Official page →

Deepgram · closed · cost: low

Pick it when latency matters, such as a live voice agent. It transcribes at 541 times real time on Artificial Analysis's benchmark.

Qwen3-ASR 1.7BHugging Face →

Alibaba Qwen · open weights · cost: free weights; you pay for the hardware · 2.5M HF downloads/mo

Pick it when audio must stay in your infrastructure. Apache 2.0, best open WER on the Open ASR leaderboard.

NVIDIA Parakeet TDT 0.6B v3Hugging Face →

NVIDIA · open weights · cost: free weights; you pay for the hardware · 599K HF downloads/mo

Pick it when you self-host at very high volume. Its throughput is about seven times Qwen3-ASR's on the same leaderboard.

Watch out for

Microsoft's MAI-Transcribe-2 scored better and cheaper on Artificial Analysis, but public access terms were not clear at the time of checking.
Whisper large v3 is still the most downloaded ASR model on Hugging Face, but it is no longer competitive on accuracy (6.36 average WER on the Open ASR leaderboard).

Doing this just once? To transcribe a recording once, use ElevenLabs Scribe v2. Scribe v2 has one of the lowest word error rates on Artificial Analysis's speech-to-text benchmark from a major vendor, labels who is speaking, and runs in a web app where you upload a file and download the transcript.

Sources

Checked 17 September 2026. Models change monthly; we re-check this page when they do.

Picking the model is the easy part.

Wiring it into a workflow that runs every week, with evals, fallbacks and a cost you can predict, is the work. Fifteen minutes, no deck.

Book fifteen minutes →