koedesk/stt-bench

koedesk STT Bench — raw transcriptions (v1, Track A)

28,400 raw transcriptions: 13 hosted speech-to-text engine variants × 12 languages × 200 FLEURS utterances, produced by a preregistered benchmark run on 2026-07-05. Every engine output is published unmodified, including full raw API responses.

Conflict of interest, stated first: this benchmark is built and funded by koedesk, a voice-typing product that uses ElevenLabs Scribe v2. We do not pretend to be neutral. Instead, everything you need to distrust us and re-score the data yourself is public: the methodology was frozen in PREREGISTRATION.md before the headline run (verifiable from git history), all raw outputs are in this dataset, and the scoring code is MIT-licensed on GitHub. Where Scribe loses, the result is published unchanged — it is not #1 overall.

• Leaderboard: https://koedesk.app/benchmark/
• Methodology, scoring code, normalizers, amendments: https://github.com/guide-inc-org/koedesk-stt-bench
• Write-ups: English · 日本語 (Zenn)

Configs

transcriptions (28,400 rows)

One row per (engine variant, language, utterance).

column	type	description
engine	string	Engine variant key (matches scores.json / leaderboard)
model_id	string	Exact model identifier sent to the API
lang	string	Benchmark language code (see FLEURS mapping below)
utt_id	string	{lang}_{index:04d} — index into the FLEURS test split, dataset order
duration_sec	float	Audio duration
ref_text_raw	string	FLEURS reference transcription, untouched
ref_text_normalized	string	Reference after the preregistered normalization pipeline
hypothesis	string	Engine output text, untouched
latency_sec	float	Wall-clock API latency for this utterance
timestamp_utc	string	When the API call was made
retries	int	Retry count for this utterance
error	string/null	Final error, if any (all rows in v1 completed successfully)
params_json	string	JSON: exact request parameters
raw_response_json	string	JSON: the full, untruncated API response

cell_scores (142 rows)

One row per (engine, language) cell, computed by the frozen scoring code (make_scores.py); scores.json is also included verbatim at the repo root. Columns include corpus CER/WER, utterance means, bootstrap 95% CIs, real-time-factor percentiles, list price per audio hour, and the tie group (tie_group, 1 = best tier). The headline metric is CER for ja/th/zh and WER elsewhere (headline_metric column); wer_* is null for th (no whitespace word boundaries).

Read results through tie groups, not raw ranks. Within a tie group, pairwise differences are not statistically distinguishable (paired bootstrap, preregistered in §6). Claiming "engine A beats engine B" inside the same tier is exactly the misuse this benchmark tries to prevent.

Engines (13 hosted variants)

engine	model_id	languages
elevenlabs	scribe_v2	12
openai	gpt-4o-transcribe	12
whisper1	whisper-1	12
google_stt_v2	chirp_3	12
deepgram	nova-3	12
mai_transcribe	mai-transcribe-1.5	12
gemini	gemini-2.5-flash	12
gemini_25_pro	gemini-2.5-pro	12
gemini_31_flash_lite	gemini-3.1-flash-lite	12
gemini_31_pro	gemini-3.1-pro-preview	12
gemini_35_flash	gemini-flash-latest	12
mistral_voxtral	voxtral-mini-2602	9 (no id/th/vi)
amivoice	-a-general	1 (ja only)

Engine selection and exclusions (e.g. engines whose terms of service prohibit publishing benchmark results) are documented in the ToS audit.

Audio provenance (audio is NOT redistributed here)

Test material is the google/fleurs test split, first 200 utterances in dataset order per language (deterministic — no sampling choices). To reconstruct the exact audio: load the config below, take rows 0–199 of the test split, then apply normalize_audio.py (16-bit PCM, −6 dBFS peak — the identical file was sent to every engine).

lang	FLEURS config	lang	FLEURS config
ja	ja_jp	ko	ko_kr
en	en_us	vi	vi_vn
zh	cmn_hans_cn	id	id_id
de	de_de	th	th_th
fr	fr_fr	es	es_419
pt	pt_br	ru	ru_ru

Known artifacts (kept as measured, per the preregistration)

• gemini_25_pro / zh / zh_0192: the model leaked ~4,800 characters of English reasoning into the transcription field for one utterance. That single utterance accounts for 86.7% of the cell's error mass (CER 48.3% as measured; 6.45% excluding it). This is the engine's real behavior, not a pipeline bug, and is published as measured.
• FLEURS zh references: 13/200 reference texts contain English annotation notes from the source dataset; this affects all engines identically.
• Gemini-family latency: measured through Cloudflare AI Gateway (affects latency_sec/RTF for the five gemini* variants only; accuracy is unaffected). All other engines were called directly.
• Measurement environment: run from Vietnam with an AWS Tokyo exit node (see Amendment 6).

Limitations

This is Track A only: FLEURS is public and every commercial engine has plausibly trained on it. These numbers connect to vendor-claimed figures and the HF Open ASR Leaderboard; they do not prove real-world dictation accuracy. A planned Track B (fresh recorded audio) was cancelled before any measurement existed — see Amendment 7. The benchmark refreshes quarterly (§9 of the preregistration).

License and attribution

• This dataset (transcriptions, scores, documentation): CC BY 4.0, © 2026 Guide Inc. (koedesk).
• Reference texts (ref_text_raw, ref_text_normalized) derive from FLEURS (google/fleurs, CC BY 4.0; Conneau et al., 2022, FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech).
• Engine outputs are published for evaluation/research purposes; each provider's terms were audited before inclusion (see ToS audit).

Citation

@misc{koedesk-stt-bench-2026,
  title  = {koedesk STT Bench: a preregistered multilingual speech-to-text benchmark},
  author = {{Guide Inc. (koedesk)}},
  year   = {2026},
  url    = {https://koedesk.app/benchmark/},
  note   = {Raw data: https://huggingface.co/datasets/koedesk/stt-bench. Methodology: https://github.com/guide-inc-org/koedesk-stt-bench}
}