cstr/fastpitch-en-GGUF

FastPitch (English) + HiFi-GAN — GGUF (ggml)

GGUF / ggml conversion of nvidia/tts_en_fastpitch + nvidia/tts_hifigan for use with CrispStrobe/CrispASR.

FastPitch is a non-autoregressive parallel TTS model that generates the entire mel spectrogram in a single forward pass (no sampling, no KV cache), making it very fast. The HiFi-GAN vocoder converts the mel to 22050 Hz PCM audio.

• Text encoder: 6-layer Transformer (384-d, 1-head, post-norm, Conv1d FFN)
• Duration predictor: 2-layer Conv1d stack + linear projection
• Pitch predictor: 2-layer Conv1d stack + linear projection
• Mel decoder: 6-layer Transformer (same architecture as encoder)
• HiFi-GAN vocoder: conv_pre + 4 upsample stages (rates 8,8,2,2) with MRF resblocks + conv_post

Single speaker, English. ~60M parameters total (FastPitch + HiFi-GAN combined in one GGUF).

Released under CC-BY-4.0 (NeMo model license).

Files

File	Quant	Size	Notes
fastpitch-en-f16.gguf	F16	~230 MB	Reference quality
fastpitch-en-q8_0.gguf	Q8_0	~120 MB	Near-lossless
fastpitch-en-q4_k.gguf	Q4_K	~70 MB	Best size/quality balance

The F16 build was rebuilt (2026-08-03), after being briefly withdrawn. The file previously published here was corrupt: the converter handed a float32 array to add_tensor(raw_dtype=F16), which labels bytes rather than converting them, so it held half the weights reinterpreted as garbage (943,872 NaNs) and would not open.

Rebuilding it exposed a second issue that briefly looked fatal — every run aborted in ggml graph compute. It was two tensors: ggml's Metal binary ops require the second operand to be F32, and enc.pos_emb / dec.pos_emb are added straight to the hidden state, so converting them to F16 killed the graph. They are kept F32 now; a matmul weight may be F16, an addend may not.

The current f16 synthesises correctly and ASR-round-trips clean.

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target crispasr-cli

# 2. Download model (auto-download also works: -m auto --backend fastpitch)
hf download cstr/fastpitch-en-GGUF fastpitch-en-q8_0.gguf --local-dir .

# 3. Synthesize
./build/bin/crispasr --backend fastpitch -m fastpitch-en-q8_0.gguf \
    --tts "Hello there, how are you doing today?" \
    --tts-output hello.wav

# 4. Verify (ASR roundtrip)
./build/bin/crispasr -m models/ggml-base.en.bin -f hello.wav

Conversion

python models/convert-fastpitch-to-gguf.py \
    --hf-model nvidia/tts_en_fastpitch \
    --hf-vocoder nvidia/tts_hifigan \
    --output fastpitch-en-f16.gguf --ftype f16

Limitations

• Single speaker only (the English model has n_speakers=1)
• Character-level tokenization (no G2P phoneme conversion yet; proper ARPABET G2P would improve pronunciation of uncommon words)
• Deterministic output (no temperature/seed controls — same input always produces same output)
• 22050 Hz sample rate

Voice provenance (EU AI Act Art. 50(4))

The upstream nvidia/tts_en_fastpitch card states it is "trained on LJSpeech". LJSpeech is 13,100 clips of a single narrator — Linda Johnson, recorded 2016–17 for LibriVox — and this is the single-speaker English checkpoint (n_speakers=1). The voice you hear is one identifiable person.

CrispASR records this as speaker_identity=real_person. Output synthesized with it carries a spoken AI disclosure, because audio resembling an identifiable person is a deep fake under Art. 3(60) whether or not any cloning took place. It does not require --i-have-rights: the donor's agreement to the training is a licensing matter settled upstream, which a downstream operator cannot attest to.

Override per run with --speaker-identity, or stamp a file permanently with models/stamp-speaker-identity.py. See docs/eu-ai-act.md §6.2a.