I Replaced My AI Narrator With Drama 3. Hear Why.

I replaced my AI narrator with Fish Audio Drama 3. Hear what changed when I gave my own cloned voice more expressive direction—and see the settings and local hardware cost behind the result. Note: This video was made with an AI and scripting workflow as part of my build-in-public journey. Please let me know if you spot any inaccuracies or discrepancies so I can continue improving the process over time. CHAPTERS 00:00 Hear the same voice, two deliveries 00:12 Why I changed my narrator 01:05 What Drama 3 adds 01:58 What changed in the controlled test 03:03 The settings I kept 03:51 What the published S2 table measures 05:04 Local GB10 timing and memory 06:01 Checking the complete take 06:57 Emotion for YouTube narration 07:47 The full original approved sample 08:29 Why Drama 3 became my default Both sides of the plain/expressive audio comparison use Drama 3. This is a comparison of delivery cues with the same voice and words, not a controlled Drama 3 versus S2.1-Pro shootout. The opening uses a fresh matched preset demonstration; the detailed measurements below come from our original complete 104-word qualification take. TESTED SETTINGS • Model: fishaudio/drama-3 • Weights revision: e0a9218aa48888270f06ef1ed9ea67622c0e17d2 • Voice: my own recording, one curated 29.396-second reference with its exact transcript • Seed: 42 | Temperature: 1.0 | Top-p: 0.9 | Top-k: 30 • Expressive cues: curious opening, excited explanation, warm invitation, curious closing questions • Runtime: NVIDIA GB10, official PyTorch implementation, CPU staging, FP8 experts / BF16 remainder, CUDA graphs ORIGINAL LOCAL TEST • 41.146 seconds of audio in 64.595 seconds of synthesis • Peak allocated GPU memory: 43.708 GiB • Separate ASR check, with no script supplied as context: 104 words, zero word edits • The live Decksmith default reproduced the approved take with identical PCM audio Synthesis timing excludes model download and loading, transcription verification, and delivery encoding. This particular local run was slower than playback. Peak GPU allocation is a measured result, not a minimum memory requirement. Settings are a tested starting point, not a promise of identical output on every runtime. WHAT THE PUBLISHED S2 RESULTS MEAN Fish's report lists English word error of 1.13% for Drama 3 and 1.62% for Fish Audio S2 on the MiniMax Multilingual Testset. The S2 number is quoted from its earlier paper using its original scoring. This is not a fresh identical-harness comparison, and word error is not an emotion score. Fish's H200 / SGLang serving results also use different hardware and software from our GB10 test. OFFICIAL SOURCES Model, weights and license information: https://huggingface.co/fishaudio/drama-3 Technical report: https://arxiv.org/html/2610.18010v1 SGLang-Omni serving cookbook: https://sgl-project.github.io/sglang-... Fish Speech source: https://github.com/fishaudio/fish-speech This video's narration is AI-generated using my own voice reference. Which matters more for your narrator: a convincing performance, or the fastest possible response? #Drama3 #FishAudio #VoiceCloning Music: I Replaced My AI Narrator With Drama 3. Hear Why. — original score — an original instrumental generated locally with MiniMax Music 3 for this video.