Xiaomi-CocktailASR-1

Target-speaker ASR SpeechLLM: given a reference speaker clip and a mixed / single-speaker audio, it transcribes only the target speaker. It also supports negative-sample rejection (empty output when the reference speaker is absent) and chain-of-thought (CoT) reasoning.

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "Ease3/Xiaomi-CocktailASR-1", trust_remote_code=True, torch_dtype="bfloat16"
).cuda().eval()

# Concatenates ref + 1s silence + target internally; returns the transcription.
text = model("target.wav", "ref_speaker.wav")            # standard
text = model("target.wav", "ref_speaker.wav", cot=True)  # <think>...</think><answer>...</answer>

Install the runtime dependencies with:

pip install torch torchaudio transformers soundfile