surogate/surogate-speech-evals

Surogate Speech evaluation results

Every transcript and score behind the numbers on the Surogate Speech model cards: jackrabbit-110m-ro, jackrabbit-110m-ro-streaming and amami-357m-ro. Other systems were run through the same harness and scorer; their rows are our own runs of the public checkpoints, not their authors' numbers.

These files are here so the scores can be checked and rescored. Please do not train on them.

Headline results

FLEURS Romanian test (883 clips), Open ASR Leaderboard multilingual runner and normalizer, one RTX 5090, bf16:

System	Params	WER	RTFx
Surogate Jackrabbit 110M, CTC + 4-gram	116M	5.69%	2,531
NVIDIA Canary 1B v2	1B	5.95%	853
Surogate Jackrabbit 110M, TDT greedy	116M	7.56%	2,774
OpenAI Whisper large-v3	1.55B	8.42%	102
SpeD ParakeetRomanian 110M	110M	8.60%	2,810
NVIDIA Parakeet TDT 0.6B v3	0.6B	11.58%	2,246

Common Voice 21 Romanian test (3,929 clips, the clip list published with SpeD):

System	WER, leaderboard normalizer	WER, SpeD normalizer
Surogate Jackrabbit 110M, CTC + 4-gram	2.07%	2.19%
Surogate Jackrabbit 110M, TDT greedy	2.19%	2.30%
SpeD ParakeetRomanian 110M	3.47%	3.58%
NVIDIA Canary 1B v2	8.65%	8.87%
OpenAI Whisper large-v3	8.94%	9.16%
NVIDIA Parakeet TDT 0.6B v3	10.06%	10.19%

Our run of SpeD gives 3.58% with its own normalizer; the SpeD paper reports 3.57%.

Streaming, FLEURS Romanian test, live 32 ms packets through the release pipeline, leaderboard references and normalizer:

Jackrabbit 110M Streaming	WER
Final text (full-context re-read + 4-gram after each pause)	7.03%
Live TDT text at each endpoint	11.22%
Leaderboard offline runner (sees only the live text)	10.67%

The final text is computed a median 84 ms (p95 182 ms) after the 640 ms pause detector fires, on one RTX 5090.

Text-to-speech, MiniMax multilingual test set, Romanian (100 sentences), judged by Whisper large-v3:

System	WER
Surogate Amami 357M, mean of Doina / Tudor / Radu	2.02% (2.65 / 1.75 / 1.66)
Facebook MMS-TTS Romanian	6.53%
ElevenLabs Multilingual v2 (published, voice cloning)	1.35%
MiniMax-Speech (published, voice cloning)	2.88%

With NVIDIA Canary 1B v2 as the judge: Amami 2.96%, MMS-TTS 5.12%. Published rows come from the MiniMax-Speech paper (arXiv 2505.07916, Table 2); those systems clone the set's reference speakers while Amami reads with its own fixed voices, so only the WER column compares.

Files

runs/asr-oal-fleurs_ro/<system>/predictions.jsonl, scores.json
runs/asr-sped-cv21_ro/<system>/hypotheses.jsonl, scores.json
runs/asr-streaming-fleurs_ro/jackrabbit-110m-ro-streaming/hypotheses.jsonl, scores.json, timing.json
runs/tts-minimax_ro/hypotheses.jsonl, scores.json
runs/checks/amami-runtime-parity.json         released native runtime vs the original build
runs/checks/amami-voice-rows.json             only the three released voices remain in the model
runs/checks/jackrabbit-gguf-parity.json       GGUF (NeMo-Speech.cpp, CPU) vs .nemo, plus both GGUF transcript sets
harness/run_eval_ml_ctclm.diff, e2_sped.py, e4_judge.py
environment.json

How it was run

• FLEURS, Open ASR Leaderboard protocol:
  • nemo_asr/run_eval_ml.py and transformers/run_eval_ml.py from huggingface/open_asr_leaderboard at commit b2e6f04, unmodified, with --dataset hf-audio/open-asr-leaderboard-multilingual-datasets --config_name fleurs_ro --split test --batch_size 64 (Whisper: batch 32, language auto-detected, as the leaderboard runs it), one RTX 5090 per run, bf16.
  • wer_oal_runner in each scores.json is the runner's own number; wer_recomputed_without_compound_pairs rescores the saved predictions without the runner's compound-word step, which is how the other tables here are scored. The ranking is the same either way.
• One deviation: the Jackrabbit CTC + 4-gram row adds seven lines to the runner that switch decoding to the model's CTC head with its bundled lm-4gram-ro.nemo (beam 32, alpha 0.5, beta 2.0), placed after the runner's own decoding setup. The full change is harness/run_eval_ml_ctclm.diff.
• Common Voice 21: the clip list from SpeD-RoASR (manifests_eval/test_cv21-clean.json), audio from Common Voice 21 Romanian, each model transcribed with harness/e2_sped.py. Scored with the leaderboard's MultilingualNormalizer (ş/ţ folded to ș/ț) and with SpeD's normalization.
• Streaming: surogate-speech eval streaming from surogate-speech with the release settings: beam 64, alpha 0.55, beta 1.75, 640 ms pauses, 512-sample packets. Final hypotheses are rescored against the leaderboard's FLEURS references.
• Text-to-speech: Amami's native CPU package synthesized the 100 MiniMax Romanian sentences with each voice; harness/e4_judge.py transcribed them with Whisper large-v3 (language ro, greedy) and Canary 1B v2 and scored with the leaderboard normalizer. MMS-TTS read the same sentences.
• Checks:
  • Amami's released native runtime was compared with the original build on 331 stress sentences × 3 voices.
  • The speaker-row check shows that the released voices stay byte-identical after the unused rows were cleared.
  • The GGUF check compares Jackrabbit's native runtime on CPU with the .nemo on identical scoring.

The client code and evaluation harness are in github.com/invergent-ai/surogate-speech.

Masked local paths

• Audio paths are replaced by clip ids: fleurs_<file id>, common_voice_ro_<id>, <sentence>-v<voice>.
• The leaderboard's NeMo runner names clips by its own order, so FLEURS clips are matched back by reference text and duration. Eleven pairs of FLEURS clips share the same sentence and duration; their rows carry "id_ambiguous": true, and their scores are unaffected because the references are identical.
• No audio is included.

Data and license

• Common Voice: only clip ids and model hypotheses are included, not reference texts or audio.
• FLEURS references are from Google's FLEURS (CC-BY-4.0).
• MiniMax test sentences are from MiniMaxAI/TTS-Multilingual-Test-Set (CC-BY-SA-4.0).
• This dataset is released under CC-BY-SA-4.0.