Fluid Speech Pro β GGUF
GGUF conversions of Fluid Speech Pro by FluidVoice
(FluidVoice-Labs/fluid-speech-pro, CC-BY-4.0) for
CrispASR (-m fluid-speech-pro). The model detects the language
itself and writes punctuation and capitals. It covers English and 24 other European languages.
Fluid Speech Pro is NVIDIA Parakeet TDT 0.6B v3 plus one
small residual block between the encoder's subsampling and its first Conformer layer:
LayerNorm β Linear β depthwise Conv1d (kernel 15) β GELU β Linear. These files were converted from the reference
.nemo checkpoint, so the block is included. The upstream transformers weights and the upstream parakeet.cpp GGUF
leave it out. CrispASR's parakeet backend runs the block whenever its tensors are present.
Files
| File | Size | Notes |
|---|---|---|
fluid-speech-pro-f16.gguf |
1.26 GB | F16 |
fluid-speech-pro-q8_0.gguf |
677 MB | Q8_0 (default for -m fluid-speech-pro) |
fluid-speech-pro-q4_k.gguf |
403 MB | Q4_K: smallest, some drift on code-switched speech |
fluid-speech-pro-clean-f16.gguf |
1.26 GB | clean mode, F16 |
fluid-speech-pro-clean-q8_0.gguf |
677 MB | clean mode, Q8_0 (default for -m fluid-speech-pro-clean) |
fluid-speech-pro-clean-q4_k.gguf |
403 MB | clean mode, Q4_K |
Clean mode produces edited text instead of a verbatim record: it drops repeated words and fillers such as
"um" and "uh". Upstream it is an optional low-rank adapter. In the -clean files it was folded into the weights
at conversion, exactly as fluid_speech.py folds it. Don't use clean mode where you need an exact record.
Usage
crispasr -m fluid-speech-pro --auto-download -f audio.wav
crispasr -m fluid-speech-pro-clean --auto-download -f audio.wav # edited, non-verbatim text
crispasr -m fluid-speech-pro-q8_0.gguf -f audio.wav # a local file
Conversion: python models/convert-parakeet-to-gguf.py --nemo fluid-speech-pro.nemo --output fluid-speech-pro-f16.gguf [--clean-mode],
then crispasr-quantize <f16> <out> q8_0|q4_k.
Validation
Compared against NeMo running upstream's loader (block attached, every tensor loaded strictly, dither 0):
- F16 stage diff (
crispasr-diff parakeet): all 28 stages pass (mel, block output, 24 encoder layers, encoder output) on jfk.wav and multispeaker.wav, for both base and clean mode. Encoder output cosine is β₯ 0.99997. Run on heavy-cpu CI withtools/ci-heavy/fluid_speech_validate.py. - Transcripts: F16 and Q8_0 match NeMo on jfk.wav, multispeaker.wav and a 30 s German/English code-switched clip, for both base and clean mode. On multispeaker.wav, clean-mode Q8_0 differs from NeMo by whitespace or punctuation only (word error rate 0). Q4_K matches on the English clips but changes 3β4 words in the code-switched opening.
- Q8_0 / Q4_K stage tolerances: the quantized files fail the F16-tuned per-stage thresholds, as stock parakeet-tdt-0.6b-v3 converted the same way does. Worst-frame encoder cosine for Q8_0 is 0.933 (Fluid) vs 0.988 (v3) on jfk.wav, and 0.9946 vs 0.9957 on multispeaker.wav.
- Source checkpoint:
fluid-speech-pro.nemoat revision46027a3b, sha256f3e377b792ae5a52a3c071adb7eba5504306f2f4bdcf933ae9da2cbcab6f0449.
License and attribution
CC BY 4.0. Please credit the model as Fluid Speech Pro by FluidVoice.
- Model: FluidVoice-Labs/fluid-speech-pro (CC-BY-4.0)
- Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, both CC-BY-4.0
- Changes in this repository: format conversion to GGUF, quantization, and (for the
-cleanfiles) the upstream clean-mode adapter folded into the weights. No retraining.
- Downloads last month
- 227
8-bit
16-bit
Model tree for cstr/fluid-speech-pro-GGUF
Base model
moondream/parakeet-ultra