[Fun-ASR-MLT-Nano-2512] Official checkpoint omits CTC weights but still returns invalid timestamps

[cutebug0523]

Summary

The official ModelScope checkpoint for FunAudioLLM/Fun-ASR-MLT-Nano-2512
contains no trained CTC decoder/head weights. FunASR initializes those layers
randomly, logs missing-key warnings, and still returns structurally valid but
semantically invalid timestamps.

Text transcription itself is accurate for English, Japanese, Korean, and Thai.
The issue is specifically the timestamp contract.

Environment

• FunASR: 1.3.14
• Model: FunAudioLLM/Fun-ASR-MLT-Nano-2512
• Source: official ModelScope master checkpoint
• model.pt size: 1,971,162,898 bytes
• model.pt SHA-256:
  f7ddbef5ff8c57254e6a045f0b316501b704fe2b774f962efc6e402364e8fe37
• Device used for reproduction: CPU
• Input: clean TTS speech, approximately 7-11 seconds per sample

Checkpoint inspection

import torch

checkpoint = torch.load(
    "model.pt",
    map_location="meta",
    weights_only=False,
    mmap=True,
)
state = checkpoint["state_dict"]

print(len(state))
print(sum(key.startswith("ctc_decoder.") for key in state))
print(sum(key.startswith("ctc.") for key in state))

Output:

1261
0
0

During AutoModel(...) initialization, FunASR reports 84 missing
ctc_decoder.* keys and 2 missing ctc.* keys.

Minimal inference path

from funasr import AutoModel

model = AutoModel(
    model="FunAudioLLM/Fun-ASR-MLT-Nano-2512",
    device="cpu",
    disable_update=True,
)

result = model.generate(
    input=["clean_tts.wav"],
    cache={},
    batch_size=1,
    language="英文",
    itn=True,
)

print(result[0]["text"])
print(result[0]["timestamps"][:3])

Observed results

Transcription similarity against the exact TTS source text:

Language	Similarity
Japanese	1.0000
Korean	0.9792
English	1.0000
Thai	0.9947

The returned timestamps are not usable:

• Every returned timestamp has score=0.0.
• Japanese: one character receives a 3.36-second span.
• Korean: one character receives a 7.14-second span.
• English: the token blush receives a 6.06-second span in a 7.46-second clip.
• Thai: one token receives a 6.18-second span.
• Timestamps remain monotonic and within the audio duration, so a structural
  validation alone does not detect the failure.

This is consistent with forced alignment running against a randomly initialized
CTC head.

Checkpoint size comparison

The official ModelScope base checkpoint is substantially larger:

Checkpoint	Size
Fun-ASR-Nano-2512/model.pt	2,127,426,538 bytes
Fun-ASR-MLT-Nano-2512/model.pt	1,971,162,898 bytes

The difference is 156,263,640 bytes, which appears consistent with the missing
CTC decoder/head parameters.

Expected behavior

One of the following:

1. Publish an MLT checkpoint containing trained ctc_decoder.* and ctc.*
   weights so the documented timestamp feature works for the 31-language model.
2. If timestamps are not supported by this checkpoint, fail closed instead of
   returning timestamps from randomly initialized layers, and document the
   limitation.

Could you confirm whether a trained multilingual CTC checkpoint exists or
whether the MLT checkpoint was packaged without those weights?

[LauraGPT]

Implemented the fail-closed behavior in #3211.

Validation against the currently published checkpoints:

• Fun-ASR-MLT-Nano-2512: 1,261 tensors and no ctc_decoder.* / ctc.* tensors. FunASR now emits one explicit warning, preserves text transcription, and omits timestamps / ctc_timestamps.
• ModelScope Fun-ASR-Nano-2512: 1,347 tensors, including 84 decoder and 2 CTC-head tensors. CTC remains enabled and the same sample returned 14 text timestamps and 13 CTC timestamps.
• Native package, remote-code, and vLLM checkpoint-loading paths are covered. Missing, partial, and shape-mismatched CTC states fail closed.
• Regression suite: 22 Fun-ASR-Nano tests passed, plus Python compilation and formatting/diff checks.

This fixes the invalid timestamp contract for the current MLT artifact; it does not claim that a separately trained multilingual CTC checkpoint exists. The issue will close automatically when #3211 merges.