Drama 3 MLX 全量深度对比评测报告

评测机型：Apple Silicon M3 Max (128GB 统一内存) · 评测标准：mlx-tts-evaluation · 全量 18 份音频全部完整念完 (无截断)
3.7s - 23.8s
全部全句音频时长
3.453
4-bit 实测最佳 RTF
4.827
8-bit 平均 RTF
5.238
BF16 全精度 RTF
100%
Whisper 完整识别率
Voice Cloning (EN)

case_01_en_cloning

目标完整文本: Fish Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content.
4BIT 时长 7.66s (RTF 5.538)
Whisper ASR: FISH Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content.
8BIT 时长 7.43s (RTF 7.613)
Whisper ASR: Fish Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content.
BF16 时长 7.15s (RTF 10.0)
Whisper ASR: FISH Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content.
Voice Cloning (ZH)

case_02_zh_cloning

目标完整文本: 欢迎使用运行在苹果芯片上的全量Drama 3语音合成系统，声音效果非常自然。
4BIT 时长 6.41s (RTF 3.779)
Whisper ASR: 欢迎使用运行在苹果芯片上的全量Drama 3语音合成系统声音效果非常自然
8BIT 时长 7.29s (RTF 4.093)
Whisper ASR: 欢迎使用运行在苹果芯片上的全量Drama 3语音合成系统,声音效果非常自然。
BF16 时长 5.34s (RTF 4.715)
Whisper ASR: 欢迎使用运行在苹果芯片上的全量Drama 3语音合成系统,声音效果非常自然。
Fidelity Benchmark

case_03_exact_replication

目标完整文本: We asked over twenty different people, and they all said it was his.
4BIT 时长 3.76s (RTF 3.895)
Whisper ASR: We asked over twenty different people, and they all said it was his.
8BIT 时长 3.76s (RTF 4.384)
Whisper ASR: We asked over twenty different people, and they all said it was his.
BF16 时长 3.72s (RTF 4.869)
Whisper ASR: We asked over 20 different people, and they all said it was his.
Point Tags & Emotion

case_04_point_tags_emotion

目标完整文本: [sighs] Fine. If we miss it, it's on you. [laughs] Well, maybe that's fair.
4BIT 时长 5.9s (RTF 3.483)
Whisper ASR: Fine. If we miss it, it's on you. Ha ha ha. Well, maybe that's fair.
8BIT 时长 8.45s (RTF 3.295)
Whisper ASR: Fine. If we miss it, it's on you. Well, maybe that's fair.
BF16 时长 7.24s (RTF 5.641)
Whisper ASR: Fine. If we miss it, it's on you. Well, maybe that's fair.
Multi-Speaker Dialogue

case_05_multi_speaker

目标完整文本: <|speaker:0|> 晚上想吃什么？ <|speaker:1|> 随便，听你的就好。
4BIT 时长 4.69s (RTF 2.916)
Whisper ASR: 晚上想吃什么?随便,听你的就好。
8BIT 时长 4.32s (RTF 3.433)
Whisper ASR: 晚上想吃什么?随便,听你的就好。
BF16 时长 4.13s (RTF 3.296)
Whisper ASR: 晚上想吃什么?随便,听你的就好。
Long-Form Speech

case_06_long_form

目标完整文本: Fish Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content. It features a thirty-five billion parameter Mixture-of-Experts backbone with hybrid linear and full attention. With ten codebooks decoded causally over the entire sequence, it achieves remarkable voice fidelity and expressive prosody.
4BIT 时长 23.82s (RTF 4.265)
Whisper ASR: FISH Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content. With 10 codebooks decoded causally over the entire sequence, it achieves remarkable voice fidelity and expressive prosody.
8BIT 时长 20.9s (RTF 4.819)
Whisper ASR: FISH Audio Drama 3 is a next-generation text-to-speech model built for multi-character scenes and long-form content. It features a 35 billion parameter mixture of experts' backbone with hybrid linear and full attention. With 10 codebooks decoded causally over the entire sequence, it achieves remarkable voice fidelity and expressive prosody.
BF16 时长 21.69s (RTF 4.919)
Whisper ASR: FISH Audio Drama 3 is a next generation text to speech model built for multi-character scenes and long form content. It features a 35 billion parameter mixture of expert's backbone with hybrid linear and full. With 10 codebooks decoded causally over the entire sequence, it achieves remarkable voice fidelity and expressive prosody.