Building a Production Greek-English Speech Recognizer

Hugging Face Daily Papers Papers
speech-recognition bilingual-asr model-ensembling data-filtering ablation-studies production-engineering

Summary

This paper details the development of Sophea, a production bilingual Greek-English automatic speech recognition system, using iterative training, data filtering, and model ensembling to meet quality gates and achieve competitive benchmark results.

We report a multi-month engineering program to build Sophea, a production bilingual Greek-English automatic speech recognition system. We evaluate the system against nine production gates covering Greek and English word error rate, language identification, and hallucinations on non-speech audio. Across twenty-three training iterations and two model architectures, no training-data composition passed all nine gates simultaneously. Meeting the Greek noisy-environment target required about 1,500 steps of dense domain exposure, while preserving English language identification tolerated only about 250 steps, or about 1,250 with a rebalanced mix that reduced Greek accuracy. We describe a six-stage data pipeline in which calibrating an audio-quality filter against in-domain anchors reduced the discarded share of scored Greek audio from 98.7 percent to 10.6 percent. A pre-registered ablation isolated a hallucination defect to one training-data package. A three-model ROVER ensemble increased gate coverage from 4-7 of 9 for individual models to 9 of 9 and reduced overlapping-speech WER from 53.35 percent to 37.87 percent, a 29 percent relative improvement. A separate learned per-clip arbiter over two models is listed as sophea/asr-k1 (preview) on the public Open ASR Leaderboard, with 4.26 percent average WER across eight public English test sets, and reaches 25.88 percent WER on live Greek noisy-environment traffic. We also document five cases in which a measurement tool produced a plausible but incorrect result and seven substantial approaches that were evaluated but not shipped. No model weights or training data are released; we report methodology and quantitative results only.
Original Article
View Cached Full Text

Cached at: 09/15/26, 10:41 AM

Paper page - Building a Production Greek-English Speech Recognizer

Source: https://huggingface.co/papers/2609.13498

Abstract

A production bilingual Greek-English speech recognition system was engineered through iterative training, data filtering, ablation studies, and model ensembling to meet strict quality gates and achieve competitive benchmark results.

We report a multi-month engineering program to build Sophea, a production bilingual Greek-Englishautomatic speech recognitionsystem. We evaluate the system against nine production gates covering Greek and Englishword error rate,language identification, and hallucinations on non-speech audio. Across twenty-three training iterations and twomodel architectures, notraining-data compositionpassed all nine gates simultaneously. Meeting the Greek noisy-environment target required about 1,500 steps ofdense domain exposure, while preserving Englishlanguage identificationtolerated only about 250 steps, or about 1,250 with a rebalanced mix that reduced Greek accuracy. We describe a six-stage data pipeline in which calibrating anaudio-quality filteragainstin-domain anchorsreduced the discarded share of scored Greek audio from 98.7 percent to 10.6 percent. Apre-registered ablationisolated ahallucination defectto one training-data package. A three-modelROVER ensembleincreased gate coverage from 4-7 of 9 for individual models to 9 of 9 and reducedoverlapping-speech WERfrom 53.35 percent to 37.87 percent, a 29 percent relative improvement. A separatelearned per-clip arbiterover two models is listed as sophea/asr-k1 (preview) on the publicOpen ASR Leaderboard, with 4.26 percent average WER across eight public English test sets, and reaches 25.88 percent WER on live Greek noisy-environment traffic. We also document five cases in which a measurement tool produced a plausible but incorrect result and seven substantial approaches that were evaluated but not shipped. No model weights or training data are released; we report methodology and quantitative results only.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2609\.13498

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.13498 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.13498 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.13498 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German

arXiv cs.CL

This paper presents a benchmark evaluating five commercial ASR systems on code-switching speech across Arabic-English, Persian-English, and German-English pairs, using a two-stage pipeline to select 300 samples per pair and assessing performance with WER and BERTScore. ElevenLabs Scribe v2 achieves the lowest overall WER (13.2%) and highest BERTScore (0.936), with public dataset available.

DialectS2S: End-to-End Speech Dialogue Modeling for Low-Resource Chinese Dialects

arXiv cs.CL

DialectS2S is an end-to-end speech dialogue model for low-resource Chinese dialects, introducing a scalable data synthesis pipeline and a two-stage post-training strategy with self-aligned speech supervision. Experiments show improvements in dialect consistency, response quality, and intelligibility, with fully open-sourced models, data, and code.