b4ph/nemotron-3-5-egg-argument-eval

ANTIMEME SYSTEMWORKS // CREATIVE UTILITIES DIVISION // DOC: AM-RS-2026 // V1.1_BUILD

NEMOTRON 3.5 TRANSCRIBE — EIGHT SPEAKERS, EIGHT LANGUAGES

I ran NVIDIA's Nemotron 3.5 Transcribe on 45 seconds of eight synthetic voices arguing about eggs, one speaker per language: English, Spanish, German, French, Portuguese, Hindi, Japanese and Mandarin. They interrupt and switch turns mid-sentence, and for 0.31 seconds three of them speak at once. Clean two-speaker demos do not test this. The question was which layer survives: words, language, script, or speaker.

The model returned 103 timestamped words and six speaker identities. Not one character of Japanese, Mandarin or Hindi script, all three of which the model card lists as supported for multispeaker ASR.

Speaker-attributed character error on this clip is 0.76 to 0.85 depending on how the prompt is worded and which GPU runs it. Diarization error is 0.4945 and does not move.

The reference is provisional, and it does not have to be. The audio was synthesized, so the generation script and the per-speaker voice stems exist in the operator's files. With those, the 25 flagged utterances become exact text, the overlap windows become exact spans, and the timestamp check becomes measurable. Hand them over and this becomes a measurement rather than agreement with an unverified annotation.

Interactive version: b4ph/nemotron-3-5-egg-argument

01 // RESULTS

CONDITION	PROMPT	cpCER	DER	JER	OVERLAP F1	WORDS	IDENTITIES
C1	operator, 8 speaker tags	0.7859	0.4945	0.5206	0.5044	103	6
C2	NVIDIA card prompt, 4 tags	0.8347	0.4945	0.5206	0.5044	116	6

Those are point values from one configuration: one wording per condition, one H100. They are not the result. Error rates are lower-is-better, overlap F1 higher-is-better.

Neither prompt is better, and the ranking does not survive replication. Rerun on an H200 the same two prompts gave 0.8157 and 0.7886, reversing the order. Four paraphrases of each prompt spread 5.7 to 9.2 points, wider than the 4.88 point gap. The machine alone moves the score too: C2's payload was preserved exactly, and its result still shifted 4.61 points between the two GPUs. Hardware and wording each move cpCER about as much as the gap, so one run per prompt carries no ranking information. Numbers, prompts and raw responses are in rerun/REPLICATION.md.

cpCER is character error with speaker attribution: each speaker's output is compared with the reference speaker it best matches, so correct words under the wrong identity are still charged. Characters rather than words because half the reference is not word-delimited. The formal definition is in evaluation/SCIENTIFIC_EVALUATION.md.

02 // WHAT THE MODEL ACTUALLY DID

Hindi came out correct and in the wrong alphabet. The line मेरी भुर्जी कभी सूखी नहीं होती was returned as meri burji kabhi suki nahi hoti. Every word is right, no Devanagari character was produced, and the score treats that as failure because the output is unusable as Hindi text.

Mandarin and Japanese collapsed into approximation. 茶叶蛋 appeared as Selene Chayote que el Rich-. The sounds are close, the language is not identified, and script compliance is zero for both across the whole recording. That is not a gap in coverage: NVIDIA's matrix lists Japanese as production-ready and Mandarin as supported, both with multispeaker ASR ticked.

Speakers merged. One predicted identity is the majority cover for four reference speakers (S03, S04, S06, S07), and a single S04 utterance carries four different tags inside five seconds.

The support claim rests on one table, so here it is. NVIDIA's card for this checkpoint, section "Language Support Matrix", at revision 9d8cd6b7:

LANGUAGE	ASR	TRANSLATION	MULTISPEAKER ASR	CODE-SWITCHING	QUALITY LEVEL
Hindi	✓	✓	✓	✓	Leading
Japanese	✓	✓	✓	✓	Production-ready
Chinese Mandarin	✓	✓	✓	✓	Supported

A different NVIDIA model, the 0.6B streaming ASR, publishes its own matrix with different tier names ("transcription-ready", "broad-coverage"). Nothing in this card relies on that one.

Overlap landed in the wrong place. The reference has one three-way window, 14.30 to 14.61 s. The model placed its three-way moments at 36.94 to 37.16 s and 37.20 to 37.22 s, and claims four concurrent voices where the reference has three. Duration overlap F1 is 0.5044 at an overlap prevalence of 19.6 %.

Even English moves. 5 to 10 errors across 24 reference words depending on wording, 0.2083 to 0.4167 word error across the ten valid runs, against 0.2917 in the published configuration. Every other language is worse, and some are too small to quote as rates at all: Portuguese has 3 reference words, French 6. The counts are in evaluation/evaluation_results.csv.

One prompt in ten broke the model. A paraphrase of the C1 instruction sent it into a repetition loop: 57,345 characters, 8,177 words, the 16,384-token limit after 42 seconds, and the timestamp middleware returned alignment_failed with zero words. That is 1 of 10 paraphrases, 1 of 11 requests. It is excluded from the cpCER statistics because there are no timestamps to score, and left in the table as the failure it was. The exclusion flatters C1: its range comes from four of five wordings, and the missing one is the worst case.

03 // METHOD

45.000 s, 44.1 kHz mono PCM, synthesized for this test. Inference: vLLM 0.32.0 with NeMo Speech on one H100 80GB, BF16 weights, MTP speculative decoding, temperature 0, audio converted to 16 kHz mono and sent as a single request. Warm inference took 4.1 s (C1) and 4.9 s (C2).

C1 asked for verbatim multilingual transcription with eight speaker tags. C2 used NVIDIA's model-card timestamp demonstration prompt, which asks for four tags. Neither reproduces NVIDIA's published evaluation protocol. Diarization error is pyannote.metrics at collar 0.

04 // LIMITATIONS

The reference is provisional: no listening pass, no native-speaker review of the Japanese, Mandarin or Hindi text, and 25 of 31 utterances flagged in evaluation/annotation_review.csv. These figures are agreement with that reference, not verified accuracy. One 45-second recording is a case study, not an estimate of the model's general quality. C1's request was never captured, but it is recoverable: the prompt and parameters come from the script that built it, and the 16 kHz audio survives with a hash matching the run record, so it can be replayed from rerun/. C2's audio bytes were redacted before storage, so its stored payload is a record rather than a replayable request.

05 // REPRODUCTION

cd evaluation
python -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python evaluate.py     # regenerates every table and metric
.venv/bin/python verify.py       # 85 checks

evaluation/ is the frozen archive nemotron_eval_final.zip (SHA-256 11318189f08d8d312d2b87a93e5d53d66833c8cf2e0356823d8b6af53e21a613) unpacked and unmodified, with that same archive available at the repository root. The report renders the installed regex version, so its hash changes with that dependency while the measurements do not.

FILE	CONTENTS
evaluation/inputs/c1_response.json, c2_response.json	raw model output, unmodified
evaluation/inputs/egg_argument_mix_44k.wav	the recording
evaluation/reference_annotations.json, annotation_review.csv	provisional reference and its review sheet
evaluation/results.json, evaluation_results.csv	every metric, both conditions
evaluation/SCIENTIFIC_EVALUATION.md	full report and scoring definitions
evaluation/evaluate.py, verify.py	scorer and 85 independent checks
evaluation/reproducibility_manifest.json	hashes, environment, missing evidence
rerun/REPLICATION.md	H200 replication, prompt paraphrases, hardware comparison
rerun/prompts.json, rerun/responses/	eleven prompts and their raw responses
rerun/c1_request_recovered.json, rerun/egg_argument_16k.wav	recovered C1 request and its exact audio

Model: nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. No weights are redistributed here. No license is granted for this evaluation material; see LICENSE_PROVENANCE.md.

---

ANTIMEME SYSTEMWORKS // CREATIVE UTILITIES DIVISION // AM-RS-2026 // V1.1_BUILD