Greek text output

[gipt66]

Hello I have some questions about the greek language recognition

1. Is romanized output the intended behavior for Greek (and possibly other non-CJK languages) with Fun-ASR-Nano-2512, or is this a bug?
2. What is the correct value to pass for language= for Greek? Is there a canonical mapping somewhere I missed (ISO 639-1/3, Chinese name, model-specific
   code)?
3. Are the miss key in ckpt: ctc_decoder.* warnings expected, or do they indicate a broken checkpoint?
4. If romanized output is by design, is there a config flag or post-processing step that recovers native script?

thank you

[LauraGPT]

Hi @gipt66, thanks for the detailed questions. Let me address each:

1. Romanized output for Greek

Greek IS one of the 31 supported languages. The romanized output likely happens because the language parameter was not set or used an unrecognized value. The model needs an explicit language hint to output native script.

2. Correct language value for Greek

The prompt template is in Chinese (语音转写成{language}), so pass the Chinese name:

from funasr import AutoModel

model = AutoModel(
    model="FunAudioLLM/Fun-ASR-Nano-2512",
    hub="hf",
    trust_remote_code=True,
    vad_model="fsmn-vad",
    vad_kwargs={"max_single_segment_time": 30000},
    device="cuda"
)
result = model.generate(input="greek_audio.wav", language="希腊语", batch_size=1)

The Chinese name for Greek is 希腊语. You can also try Greek — the underlying Qwen3 LLM may understand English language names too, but the Chinese form is more reliable.

3. ctc_decoder.* missing key warnings

These are expected and harmless. The CTC decoder weights exist in the checkpoint for the vLLM inference engine path but are not loaded when using standard AutoModel inference. Safe to ignore.

4. Native script recovery

Once you pass language="希腊语", the model should output native Greek script directly — no post-processing needed. If it still romanizes with the language hint, please share the specific audio or output so we can investigate further.

[LauraGPT]

Correction to my earlier reply: Greek is not supported by the base FunAudioLLM/Fun-ASR-Nano-2512 checkpoint. The base model covers Chinese (including Cantonese), English, and Japanese. Greek is supported by FunAudioLLM/Fun-ASR-MLT-Nano-2512.

from funasr import AutoModel

model = AutoModel(
    model="FunAudioLLM/Fun-ASR-MLT-Nano-2512",
    trust_remote_code=True,
)
result = model.generate(
    input="your_greek_audio.wav",
    language="希腊语",
    llm_kwargs={"do_sample": False},
)

The romanized output from the base checkpoint is therefore a model/language mismatch, not a Greek post-processing mode. I also found contradictory language claims in the base Hugging Face model card; a bilingual correction is prepared.

If miss key in ckpt: ctc_decoder.* still appears with the current MLT checkpoint, please share the FunASR version and complete model-loading log. That warning alone does not establish that the checkpoint is broken.