Transcribe for 83% less than GPT-Transcribe

Turn recorded calls, meetings and interviews into searchable text. SIE runs NVIDIA’s open Parakeet speech model behind one API call: send a recording, get its transcript.

Get started Run in your cloud

Parakeet TDT 0.6B v3 on SIE

Figures kept

Roughly 60% of the CFM fleet is parked today. In line with this, we're also seeing significant headwinds …

Transcripts come back with the wrong names and numbers, or cut short without an error

OpenAI GPT-Transcribe · English, temperature 0

Same error in both runs

A figure
Earnings call · 13 s

Said

The 64 million growth of other …

Returned

The 60 million złoty growth of other …
A country
Earnings call · 7 s

Said

In Sweden Consumer, we expect the momentum …

Returned

As we in consumer, we expect the momentum …
A retailer
Design meeting · 8 s

Said

… chain companies like Carphone Warehouse

Returned

… Chain companies like Carrefour and Warehouse.
A currency
Design meeting · 4 s

Said

And so we've got we've still got four euros to go to spend.

Returned

And so we've got, we've still got four URLs to go to spend.

OpenAI GPT-4o Mini Transcribe · default, no chunking

Cut short, HTTP 200

111 of 219 recordings, none longer than 12 minutes, came back with part of the transcript missing, and no error.

Each reply stopped at the model’s 2,048-token output limit. On 20 of 44 earnings calls, a minute or more of speech is gone.

End of one reply · Earthstone Energy earnings call, third 12-minute piece

 HTTP 200 "text": "… related to the fourth quarter earnings and full year, we had a one-r", "usage": { "output_tokens": 2048, … }

OpenAI Developer Community Gpt-4o-transcribe truncates output after ~8-9 minutes even on short segments APIBugs tomfordrumm Aug 26, 2025
“no error message, just a truncated transcription.”
3 replies

arXiv:2512.04118 UC Berkeley and UCSF Clinicians’ safety reports on an AI scribe: 18.5% cite wrong drug names, doses or instructions
Patient Safety Risks from AI Scribes: Signals from End-User Feedback

Dai, Huang, Nasrallah and seven others

“incorrectly recorded medication names, dosages, and instructions (18.5% of safety-related feedback)”
1 December 2025

Accuracy comparable to GPT-Transcribe on 39 hours of earnings calls

How we tested

Word error rate

39 h of calls ↓

Parakeet TDT 0.6B v3

Parakeet TDT 0.6B v3 8.1%

GPT-Transcribe

GPT-Transcribe 9.2%

GPT-4o Mini Transcribe

GPT-4o Mini Transcribe 9.9%

Price

per 1,000 audio hours ↓

Parakeet TDT 0.6B v3 on SIE

$45 Parakeet TDT 0.6B v3 on SIE

GPT-4o Mini Transcribe

$99 GPT-4o Mini Transcribe

GPT-Transcribe

GPT-Transcribe $270

Earnings call evaluation

44 calls, 39.3 hours of audio

Audio: Earnings-21 calls, the same 12-minute pieces for every model
Word errors: Every wrong, missing or extra word. Parakeet’s rate is lowest.
Content words: The pre-registered primary measure skips words like “the”. There GPT-Transcribe is lower: 7.8% against 8.9%.
Prices: SIE’s Serverless rate and OpenAI’s list prices, 11 Oct 2026. GPT-4o Mini Transcribe ran with chunking on and is priced from its tokens.
Claude: Anthropic’s models take no audio input

Send a recording, get its transcript

View on GitHub

from pathlib import Path
from sie_sdk import SIEClient
client = SIEClient(api_key="API keysk-sie-…")
audio = {"data": Path(fileWAVami-es2004b-carphone-warehouse.wavbrowse).read_bytes(), "format": "wav"}
result = client.extract(
    "modelnvidia/parakeet-tdt-0.6b-v3",
    {"audio": audio},
)
print(result["data"])

audio_bytes=$(base64 < 'ami-es2004b-carphone-warehouse.wav' | tr -d '\n')
curl https://api.superlinked.com/v1/extract/nvidia%2Fparakeet-tdt-0.6b-v3 \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"audio\":{\"data\":\"$audio_bytes\",\"format\":\"wav\"}}]}"

Build the "Speech to text" capability into my app using the Superlinked Inference Engine (SIE).

Context
- SIE is an OpenAI-style inference API. Python SDK: `from sie_sdk import SIEClient`; TypeScript: `@superlinked/sie-sdk`.
- Base URL: https://api.superlinked.com (or my regional endpoint). Auth: Bearer key from env `SIE_API_KEY` (never hard-code it).
- Model: nvidia/parakeet-tdt-0.6b-v3 (SIE primitive: /extract). Keep the model id configurable.

Task
- Input: an uploaded audio recording.
- Behaviour: return a verbatim transcript of the uploaded audio
- Send one POST /v1/extract request (SDK: extract) with one item {audio: {data, format}}: the original audio bytes (base64 in a JSON request) and the file extension as format. Send no params. Do not add an instruction: Parakeet takes none and rejects a prompt. Read the transcript from the returned item's data.text.

Deliverables
- A typed client wrapper, an application-level function for this task, error handling for timeouts/empty input, and unit tests with a stubbed client.
- Wire it into my existing stack (ask me which framework if unclear) and add a short usage example.

Run

Output

data.text

That's kind of what led us on to comparing TV remote controls with their design features. Chain companies like Carphone Warehouse.

data.duration_ms: 8340

Transcribe on Serverless or in your own cloud

Serverless API

We host the models; you call one API.

No idle GPUs, pay for what you use
SOC 2 Type 2, US or EU data residency
$50 of credit to start, no card required

Get started

Self-hosted (Kubernetes)

SIE in your clusters, backed by our engineers.

Data stays in your own cloud account
Same models and API as Serverless
Terraform to your cloud in minutes

Talk to us Deploy guide

Run locally

The open-source server on your own machine.

docker run --gpus all -p 8080:8080 ghcr.io/superlinked/sie-server:latest-cuda12-transformers5

Quickstart

Switch to SIE in 5 minutes

from openai import OpenAI

client = OpenAI()
result = client.audio.transcriptions.create(
    model="gpt-transcribe",
    file=open("call.wav", "rb"),
)
print(result.text)

SIE

from openai import OpenAI

client = OpenAI(api_key="sk-sie-…", base_url="https://api.superlinked.com/v1")
result = client.audio.transcriptions.create(
    model="nvidia/parakeet-tdt-0.6b-v3",
    file=open("call.wav", "rb"),
)
print(result.text)

Transcripts come back with the wrong names and numbers, or cut short without an error

A figure

A country

A retailer

A currency

Accuracy comparable to GPT-Transcribe on 39 hours of earnings calls

Send a recording, get its transcript

Transcribe on Serverless or in your own cloud

Switch to SIE in 5 minutes