Eng-Ahmd/Drama-3-Windows-GGUF

Drama 3 · Windows GGUF

Expressive speech, voice cloning and contextual editing on your own hardware.

A Q4_K / Q8_0 release of Fish Audio Drama 3, paired with a Windows inference engine that places routed experts across system RAM and CUDA. Built with Fish Audio.

Download the model · Windows runtime · Installation

• 35.9B MoE — 256 experts per layer, eight selected per token.
• Expressive voice cloning — reference-conditioned speech with emotion and event tags.
• Audio-history conditioning — contextual edits and continuity between long-text segments.
• Local CPU/CUDA execution — configurable expert placement for limited GPU memory.
• 44.1 kHz audio — 21.67 GiB model and 1.46 GiB codec.

Listen: original vs Windows

The three demos from the original model card, reproduced with the same voice reference, requested text and prior-take context. Original recordings are provided by Fish Audio; the Windows recordings below were generated with this quantized release.

Shared voice reference

The original en0 reference clip used by all three examples:

1 · Expressive point tags

A single voice moves from a sigh to spoken dialogue and laughter, using [sighs] and [laughs] in the prompt.

Fish Audio original	Windows · Q4_K / Q8_0

2 · Edit with context

Change Wednesday → Thursday in a spoken report reminder. The earlier take is supplied as acoustic history, preserving the voice and context while revising the line.

Fish Audio original	Windows · Q4_K / Q8_0

3 · Chained edit

Change the deadline from noon → five, with both earlier takes retained as acoustic history.

Fish Audio original	Windows · Q4_K / Q8_0

Demo settings and reproduction

Local recordings use seed 123, temperature 0.7, top-p 0.7, top-k 30 and twelve GPU expert layers. The source demos do not publish their random seeds or sampling settings. Generation is stochastic; timing, expression and nonverbal events can differ.

The contextual examples include generated acoustic tokens from the earlier takes, not just their transcripts. After installing the runtime, download the shared reference and generate the demos:

Invoke-WebRequest -Uri "https://github.com/user-attachments/files/33393249/en0.wav" -OutFile en0.wav
python -m fish_drama3.watchdog --output runs/demo-watchdog -- python -m examples.generate_model_card_demos --bundle models/drama3-windows/model --codec models/drama3-windows/codec/codec.gguf --reference en0.wav --output runs/demos

Run the demo generator through the runtime watchdog to reproduce the sequence. Demo settings and measurements.

Podcast · The Small Signal

A complete 592-word podcast in one generation request. This 3 minute 16 second recording uses the full script at once: no text chunks, retries, selected takes or speech edits.

Press play to listen here. This player pairs the full original narration with static artwork; it does not add speech edits. The original WAV and MP3 are available below.

WAV · MP3 · Full script · Importable settings and provenance · Quality checks

Generated through the 0.2.0 browser studio with the en0 reference above, seed 123, temperature 0.7, top-p 0.7, top-k 30 and a 6144-frame budget. Automatic placement selected 15 GPU expert layers with a 3% VRAM reserve; the cache context was 10,240 rows. The model ended normally at 4216 frames / 195.79 seconds. The codec decoded the continuous acoustic stream in bounded context windows (1792 frames plus 256 context), without editing speech.

On Windows 11 with an RTX 5080 Laptop (16 GiB VRAM), 128 GiB RAM and 8 CPU threads, acoustic generation including prefill took 520.00 seconds; CPU codec decoding added 223.34 seconds, for 743.34 seconds synthesis. These exclude model loading (27.26 seconds) and MP3 export. This is one recorded run, not an aggregate speed benchmark.

The untouched WAV has 0.34% ASR WER (2 edits / 592 reference words), no full-scale clipped samples, peak magnitude 0.54089 and a longest measured near-silent interval of 0.06 seconds. Automated transcription is approximate and does not replace listening or measure speaker similarity. This single example establishes that this full script completed in one request; other text, seeds and environments can still omit or repeat words, stop early or pause.

To try it, load The Small Signal in the studio, upload the shared en0 reference and import the linked settings. Public settings use standard download paths; adjust them for your installation. Automatic expert placement and seeded output can differ with available memory, hardware and runtime.

Model & runtime

Component	Configuration
Speech model	35.9B MoE, hybrid full / linear attention
Routed experts	Q4_K · 8 of 256 selected per token
Backbone	Q8_0; sensitive vectors in F32
Audio tower	BF16 · 10 acoustic codebooks
Codec	Native reference encoder / decoder; F32 learned weights
Output	Mono 44.1 kHz PCM16 WAV
Download size	Model 21.67 GiB · codec 1.46 GiB
Engine	Windows runtime 0.1.1 or newer

Execution path: PyTorch runs the backbone and audio tower; GGML CPU/CUDA runs routed experts; the native codec encodes the voice reference and decodes the generated audio. The fish_drama3 GGUF files require this matching engine.

Measured generation

Windows 11 · RTX 5080 Laptop, 16 GiB VRAM · 128 GiB RAM · 12 GPU expert layers · 8 CPU threads. These are warm-model, single-request measurements for the recordings above, with CPU codec processing.

Example	Audio length	Generation	Gen + decode	RTF
Expressive tags	7.85 s	40.30 s	47.87 s	6.10
Contextual edit	4.27 s	13.38 s	17.28 s	4.04
Chained edit	4.37 s	14.35 s	18.32 s	4.20

RTF is synthesis time divided by audio length. Timings include model prefill, generation and native codec decode; they exclude one-time model loading, reference encoding and file writing. Model loading took 23.04 seconds for this run.

Generation is offline. The external CLI watchdog reserves at least 1 GiB GPU memory and 8 GiB host RAM; expert placement is configurable. Prebuilt CUDA libraries target SM120. The documented environment uses Python 3.13, PyTorch 2.13.0+cu132, CUDA Toolkit 13.3 and the Visual C++ x64 runtime.

Browser generation app

The Windows Studio loads and unloads the model, provides the full podcast script, exposes supported sampling and placement settings, and saves WAV or MP3 with settings JSON. Runtime 0.2.1 adds GPU audio-tower CUDA graphs, expert bridge optimizations, live timing counters and speech-budget preflight. Install the matching source and native release.

Single pass submits the whole script in one request, with up to 8,192 frames inside a 10,240-row prompt-plus-output context. Reference audio consumes context too. Smart segments preserve words and tags, prefer sentence boundaries and share up to 65,536 total frames across at most 32 passages. The 32–2000-character passage bound is an application limit. Preflight estimates speech length and warns before a likely budget overrun; estimates do not guarantee complete or accurate speech. Generation percentages are labeled estimates; frame-budget usage and codec-window counts are measured separately.

Automatic placement reserves graph buffers and request caches first, then fits expert banks using actual free VRAM, a 256 MiB workspace allowance and the selected 3% VRAM reserve. Manual mode supports 0–40 expert layers; fit depends on hardware. The host monitor preserves 8 GiB free RAM. VRAM occupancy and GPU compute utilization are separate measurements.

A matched warm 64-frame test on an RTX 5080 Laptop measured approximately 28% higher acoustic decode throughput with tower graphs. This excludes cold loading and is not a complete podcast speed claim. See the performance notes and reproduction scripts for conditions and CPU-policy tradeoffs. The existing CPU codec and OpenMP backend remain enabled because alternatives were slower in the tested workloads.

MP3 export requires FFmpeg. Start the installed source with python -m fish_drama3.webui, then open http://127.0.0.1:7860.

Download & run

Install the Windows runtime and Python environment, then download the model and codec with their manifests:

python -m pip install "huggingface_hub>=1.0,<2"
python -c "from huggingface_hub import snapshot_download; snapshot_download('Eng-Ahmd/Drama-3-Windows-GGUF', local_dir='models/drama3-windows', allow_patterns=['model/*', 'codec/*', 'LICENSE', 'LICENSES/*', 'NOTICE.md'])"
fish-drama3 synth --bundle models/drama3-windows/model --codec models/drama3-windows/codec/codec.gguf --text "Every voice has a story. Let's make yours heard." --output outputs/hello.wav --gpu-expert-layers 12

For voice cloning, add --reference-audio and its --reference-text. Long text uses bounded segments with audio-history conditioning. Complete output requires EOS; the CLI verifies file hashes and preserves existing results.

The root-level Drama-3-35.9B-Q4_K-Q8_0.gguf is an identical alias of model/model.gguf; download one copy of the model.

Quality & validation

• Reference codec: all 850 checked acoustic codes matched; decode relative RMSE 1.74 × 10⁻⁶ against the pinned reference.
• Intelligibility: four short English checks had 0.0 ASR WER; a Chinese plain-speech check had 0.0 CER.
• Runtime stability: a guarded two-hour run completed 713 speech requests and 143 validation recovery cycles, with identical acoustic codes for the repeated seeded request.

These are targeted checks, not aggregate speech-quality or speaker-similarity scores. Seed-dependent incomplete generation and the evaluation method are documented in the quality report.

Credits & license

Model and reference implementation: Fish Audio Drama 3 and Fish Speech. Native codec integration derives from s2.cpp; backend and quantization components use GGML. audio.cpp and Jmica's Drama 3 GGUF informed the porting survey. Pinned revisions and component attribution are in NOTICE.md.

Model-derived weights retain the Fish Audio Community License v1.0. The donor-derived codec is separately governed by the Fish Audio Research License, which requires a written Fish Audio agreement for commercial use. GGML components retain their MIT license. See LICENSE and component licenses.

Use voice references with the necessary rights and speaker consent. This is an independent Windows integration. Built with Fish Audio.