Status: Preview. Sophea ASR K1 is served as a hosted API. Self-service registration is not open yet; access is granted on request (see below). Benchmark results published for this model are marked as a preview until self-service registration is available.
To evaluate Sophea ASR K1, send an email to info@kiefer.gr or info@sophea.ai and we will send you an API key. Evaluation keys are provided free of charge to researchers and leaderboard maintainers; pilot pricing for production use is available on request.
Or copy this message:
To: info@kiefer.gr Cc: info@sophea.ai Subject: Sophea ASR K1 — API key request Hello KIEFER team, I would like to request an API key to test Sophea ASR K1. Name: Organisation: Intended use (e.g. evaluation, benchmark, pilot): Languages (en / el): Approx. audio volume: Thank you,
We usually reply within one business day. Keys are personal; please do not share them publicly.
Sophea ASR K1 is KIEFER SA's production speech-to-text system for English and Greek (incl. Greek/English code-switching): a two-model, per-clip-arbitrated system built on a far-field-repaired 1.7B encoder-decoder ASR model and a 1B AED model, served on a single GPU. It targets meetings, calls and far-field microphones (AMI-cleaned WER 7.28%) as well as read and broadcast speech (LibriSpeech test-clean 1.16% / test-other 2.66%).
POST {API_URL}/v1/transcribe — multipart/form-data. The base URL {API_URL} is sent to you together with your API key; it is not published here.
| field | type | notes |
|---|---|---|
file | audio file | wav / flac / ogg / mp3, any sample rate (resampled to 16 kHz mono); ≤ 60 s per request (long-form: segment client-side) |
language | string | en (default) or el |
model | string | asr-k1 (default) |
Header: X-API-Key: <your key>. Response: {"text": "...", "system": "sophea-asr-k1", "duration_s": 12.3, "latency_s": 0.41}. Health check: GET {API_URL}/health.
curl -s -X POST $API_URL/v1/transcribe -H "X-API-Key: $KEY" -F file=@clip.wav -F language=en
import requests
r = requests.post(f"{API_URL}/v1/transcribe", headers={"X-API-Key": KEY}, files={"file": open("clip.wav","rb")}, data={"language": "en"})
print(r.json()["text"])
Batch throughput: ~100–120× real time per GPU with 16–32 concurrent requests. Access: self-service registration is not open yet — the endpoint URL, API keys and pilot pricing are provided on request: info@kiefer.gr / info@sophea.ai.
Measured 2026-09-04 with the leaderboard's own api/run_eval.py on current main (8 public English datasets, including the new Earnings22-Cleaned-AA-chunked and Monsoon en_IN sets), English forced, through this API. Word error rate in %.
| AMI-Cleaned | Earnings22-Cleaned-AA-chunked | Gigaspeech-Cleaned | LS Clean | LS Other | SPGISpeech | Voxpopuli-Cleaned-AA | Monsoon en_IN | Avg (8) |
|---|---|---|---|---|---|---|---|---|
| 7.28 | 6.09 | 7.65 | 1.16 | 2.66 | 2.72 | 2.90 | 3.65 | 4.26 |
Independent run by the Open ASR Leaderboard maintainers through the same API (August 2026, 7 datasets without Monsoon): 4.35 average WER, matching the per-dataset numbers above (see PR #201). Results are listed as a preview while self-service registration is pending.