ga642381/TiCo-StepAudio2-TiCoBench-v2-outputs

Step-Audio-2-mini outputs on TiCo-Bench-v2

Generated speech for all 2,000 TiCo-Bench-v2 items, from two models: the base Step-Audio-2-mini and the TiCo version trained on it. Released so the duration numbers in the paper can be recomputed from the audio rather than taken on trust — every figure below is derived from these files and nothing else.

The two runs

directory	model
vanilla/	stepfun-ai/Step-Audio-2-mini, unmodified
tico-grpo-ckpt5400/	the same base, SFT on Spoken Time Markers then GRPO with CHORD on a duration reward, merged at step 5400

TiCo's recipe: base → SFT adapter (checkpoint-625) → GRPO+CHORD with rewards duration, time_presence, monotonicity, repetition_penalty, copy_penalty, ratio_consistency → merged.

Results

Spoken duration against the instructed duration. MAPE divides by the instructed duration; mean is signed, so negative means the model spoke for less time than it was asked to.

	n	MAE	MAPE	mean
vanilla	2,000	19.01s	65.6%	−6.93s
TiCo	2,000	7.19s	22.4%	−2.28s

Per subset, MAPE:

subset	n	vanilla	TiCo	Δ
Creative	200	68.0%	19.1%	−48.9
QA	1,000	61.1%	20.1%	−41.0
Reasoning	600	62.5%	23.6%	−38.9
Summarization	200	94.8%	33.5%	−61.3
ALL	2,000	65.6%	22.4%	−43.2

Summarization is the largest gain, not the smallest. It is also the subset where both models overshoot (vanilla +13.72s, TiCo +5.02s) while undershooting everywhere else — a property of the base model on long input audio that TiCo reduces by about two thirds.

Two caveats that matter for reading the numbers

A generation cap affects the tail. Decoding used max_tokens 2048, which under Step-Audio-2's text/audio interleave is roughly 65 s of speech. Responses that hit it stop there regardless of the target. The cap applies identically to both models, so the comparison is sound, but the rate differs and it is concentrated in Summarization:

at the 2048-token cap	vanilla	TiCo
Creative	14.5%	9.5%
QA	10.3%	7.5%
Reasoning	1.3%	6.8%
Summarization	37.0%	17.5%

On the Summarization items that finished on their own, vanilla scores 14.05s / 61.0% (n=126) and TiCo 5.80s / 22.2% (n=165).

Reasoning looks backwards — vanilla reaches the cap far less often — because vanilla undershoots Reasoning by 18.98s on average. It stops early and largely ignores the instruction, so it rarely gets near the ceiling. Reaching the cap more often is a side effect of trying to fill the requested duration.

Summarization needed a memory fix to run at all. Its XSum input audio is 36–147 s where every other subset medians 5–9 s, and the audio encoder allocates (B, 20, T, T) then upcasts it, so one long item needs about 9 GB of transient activation. vLLM does not budget for it: max_chunk_size = 29 is used only to build the dummy audio for memory profiling, so the engine sizes its KV cache believing no item exceeds 29 s. These runs used gpu-memory-utilization 0.60, max-num-seqs 1, concurrency 1; at the defaults the subset OOMs and earlier result sets held 1,785 items rather than 2,000.

Layout

vanilla/
  TiCo-Bench-<Task>-<band>_<source>.wav    generated speech
  TiCo-Bench-<Task>-<band>_<source>.txt    prompt, serialized input, and the
                                           model's text after vllm_text_output:
  distill.shard_0.jsonl                    prompt/output token ids per item
tico-grpo-ckpt5400/                        same layout
TiCo-Bench-v2/metadata.jsonl               id, input audio path, question, and
                                           `solution` = the instructed seconds

The TiCo model's .txt files contain <X.X seconds> Spoken Time Markers in the generated text; the vanilla model's do not. vLLM server logs are omitted — they are decoder debug output, not results.

Input audio is not included. It comes from TiCo-Bench-v2 and is referenced by path in metadata.jsonl.

Recomputing the table

Duration is read from the WAV header and compared with solution from the metadata — no alignment or ASR involved:

import json, wave, contextlib
meta = {json.loads(l)["id"]: json.loads(l)["solution"]
        for l in open("TiCo-Bench-v2/metadata.jsonl")}
errs = []
for sid, target in meta.items():
    with contextlib.closing(wave.open(f"tico-grpo-ckpt5400/{sid}.wav")) as w:
        spoken = w.getnframes() / w.getframerate()
    errs.append(abs(spoken - target) / target * 100)
print(sum(errs) / len(errs))     # MAPE