EmphAssess 重音检测 · 913 条 · 2026-11-10

refine checkpoint18000 Instruction 合并(1.0)

同一份权重,两种问法。问答格式下模型大多答错;把 <PCAP> 接在音频后面并预填半句话,让它自己写出重读的词,F1 从 22 升到 79。

结果

方式systemPRF1副口径 F1无答案
基线:问答格式带19.7024.9222.0126.3410
预填:直接取引号里的词带79.2979.3779.3386.420
P / R / F1
micro,913 条全部计入,原版 SpeechCombine 打分器(区分大小写)。
副口径
预测和标准答案都转小写,数字和英文数词互认("3" 算作 "three")。
无答案
没解析出任何词的条数。
system
两行最前面都带同一段 system 提示。

一条音频的例子

And so we went straight into, like, feeding baby lambs.

标准答案:重读 we。下面的 <音频> 代表这条语音的一串 CV3 音频 code。

预填

灰底 给模型的输入,到最后那个引号为止 绿底 模型接着写的
<|im_start|>system
You are a helpful assistant.
    Always follow the user's instructions precisely and respond with what is required.
    Please read the instructions carefully before answering.
    Your responses must strictly align with the user's intent and constraints.
    <|im_end|>
<|im_start|>user
And so we went straight into, like, feeding baby lambs.[SEP_1]<音频>[SEP_2]<|im_end|>
<PCAP>Based on the audio prosody, the speaker stresses the word "we" and speaks with a high level of energy, making the delivery sound excited and lively.</PCAP>

取引号里的词当答案:we对

这一行没有问句,也没有 assistant 回答那一步,答案直接从 <PCAP> 里读。

基线:问答格式

<|im_start|>system
(同一段 system 提示)<|im_end|>
<|im_start|>user
Answer the question from the user. Put your answer in the format "Emphasized word(s): "___" ". Do not repeat your answer. Audio and prosody: And so we went straight into, like, feeding baby lambs.[SEP_1]<音频>[SEP_2] Which word(s) in the audio clip are emphasized?<|im_end|>
<|im_start|>assistant
Emphasized word(s): feeding

回答:feeding错

回答之后模型还接着生成了一段语音,检测只对文字部分打分。

看这个结果要注意

结果目录 /proj/spans/lyf/tokenizer_comparison/refine_yodas_librilight_r64_checkpoint18000_instruct100_emphassess_detection_pcapelicit_20261110T195549Z/RESULTS.md 预填的逐条输入输出 /proj/spans/lyf/tokenizer_comparison/refine_yodas_librilight_r64_checkpoint18000_instruct100_emphassess_detection_pcapelicit_20261110T195549Z/full.prefill_sys.stage1_prompt_response.jsonl 基线 /proj/spans/lyf/tokenizer_comparison/refine_yodas_librilight_r64_checkpoint18000_instruct100_emphassess_detection_free_20261109T121253Z/emph_detection_prompt_response.jsonl