TADA-3B-ML — GGUF (ggml-quantised)

GGUF / ggml conversion of HumeAI/tada-3b-ml for use with CrispStrobe/CrispASR.

TADA-3B-ML is a multilingual text-to-speech model built on Meta Llama 3.2 3B with a flow-matching (FM) speech decoder and custom Hume codec. Key property: 1:1 token alignment — every text token maps to exactly one speech vector, eliminating transcript hallucination. Supports 10 languages (en, es, ja, zh, de, fr, it, pt, ko, ar). 24 kHz mono output.

License: Apache-2.0 / Llama 3.2 Community License ("Built with Llama").

Files

The Q4_K file uses a TADA-aware quantization policy (tail=14): large transformer block matrices are quantized, while the last 14 token-embedding rows and all tada.* flow-matching tensors are kept at F16. This preserves the timing and acoustic conditioning paths where quantization noise matters most.

Architecture

Text Input
  |
BPE Tokenize (Llama-3.2 128K vocab)
  |
Llama-3.2-3B AR Forward (28L, 3072d, 24 heads / 8 KV)
  + acoustic embedding (512d) + gray-code time embedding
  |
Flow-Matching Speech Head (6L AdaLN + SwiGLU, 10 Euler steps)
  |-- noise → speech vector (528d)
  |
TADA Codec Decoder (DAC upsampler)
  |-- speech vectors → 24 kHz PCM
  |
Output: float32 mono @ 24 kHz

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

# 2. Pull model + codec + default voice reference
huggingface-cli download cstr/tada-tts-3b-ml-GGUF \
  tada-tts-3b-ml-q4_k.gguf tada-codec-f16.gguf tada-ref.gguf \
  --local-dir .

# 3. Synthesize (with default voice reference)
./build/bin/crispasr --backend tada \
  -m tada-tts-3b-ml-q4_k.gguf \
  --codec-model tada-codec-f16.gguf \
  --voice tada-ref.gguf \
  --tts "Hello, this is a test of the TADA speech synthesis system." \
  --tts-output hello.wav \
  --seed 42

For near-lossless quality use tada-tts-3b-ml-q8_0.gguf; for the F16 reference quality use tada-tts-3b-ml-f16.gguf.

Recent CrispASR builds can also resolve this repo through the model registry:

./build/bin/crispasr --backend tada -m auto --auto-download \
  --tts "Hola, esto es una prueba del sistema TADA." \
  --tts-output hello.wav

Voice cloning from your own audio

Build a reference voice from any .wav with the built-in --make-ref pipeline (no Python needed). It needs the tada-encoder-*.gguf + tada-aligner-*.gguf from this repo — add --auto-download and they are fetched automatically:

# 1. Build a reference GGUF from a voice sample + its EXACT transcript
./build/bin/crispasr --backend tada-3b-ml -m tada-tts-3b-ml-f16.gguf --auto-download \
  --make-ref --voice your-voice.wav \
  --ref-text "Exact words spoken in your-voice.wav." \
  --make-ref-output my-voice.gguf

# 2. Synthesize in that voice
./build/bin/crispasr --backend tada-3b-ml \
  -m tada-tts-3b-ml-q4_k.gguf --codec-model tada-codec-f16.gguf \
  --voice my-voice.gguf \
  --tts "Text to speak in the cloned voice." --tts-output output.wav

The --ref-text must match the audio (it drives the text↔audio alignment). For non-English audio pass --language <code> to select tada-aligner-<code>.gguf. Or pass any voice reference GGUF directly via --voice /path/to/voice.gguf.

The bundled tada-ref.gguf encodes a short JFK clip as the default voice.

Forced-alignment word timestamps (--align)

The TADA aligner (a wav2vec2 CTC model over the Llama-3.2 BPE vocab) also does forced alignment: given audio + its transcript it emits frame-accurate word timings. Same assets as --make-ref (auto-downloaded):

crispasr --backend tada-3b-ml -m tada-tts-3b-ml-f16.gguf --auto-download \
  --align --voice speech.wav --ref-text "exact transcript" \
  --align-format srt   # srt (default) | json | plain

Multilingual: pass --language <code> to use tada-aligner-<code>.gguf (ar ch de en es fr it ja pl pt). Note it is a forced aligner — it needs the transcript, it is not a standalone recogniser.

Source model

• Upstream: HumeAI/tada-3b-ml (BF16 safetensors)
• Base model: meta-llama/Llama-3.2-3B
• Codec: HumeAI/tada-codec
• Paper: arXiv:2602.23068
• Converted with: models/convert-tada-to-gguf.py, models/convert-tada-codec-to-gguf.py, and crispasr-quantize
• Runtime: CrispStrobe/CrispASR

Validation

The Q4_K model was validated with CrispASR by synthesizing "Please call Stella." and ASR-roundtripping the output with Whisper tiny.en:

Please call Stella.

Runtime fixes (CrispASR ≥ commit 0a95b326)

This GGUF was produced alongside the following CrispASR runtime fixes; use a build from that commit or later:

• Voice prompt AR continuation — C++ now correctly inserts the n_prompt PAD token slots that Python uses for the voice replay phase. Without this fix, long voice references caused the AR generation loop to run 0 iterations, producing near-silence.
• BF16 noise parity — FM denoising noise is now rounded to BF16 before scaling, matching PyTorch's torch.randn().to(bfloat16) behaviour and eliminating subtle residual drift.
• Vulkan contiguity — ggml_cont() wrappers added around strided 2D views in the B2 FM graph; required for Vulkan element-wise kernels.
• Mixed-precision Q4_K — crispasr-quantize tail=14 keeps the last 14 token-embedding rows and all TADA tensors at F16, stabilising the timing path for the 3B model.

Checksums

0a738e55d88af1ed5c4106bf12bf6cf8e4f134c50a2963b75d9d94c26c10f0cf  tada-tts-3b-ml-f16.gguf
5606670f6787fbb186fc3371ac418d1e6b7fbfe879f4fc8b6c82294eb9a15efa  tada-tts-3b-ml-q4_k.gguf
23da2b8230c2e7b1f8753b80200d059c75e0752ef37626917078952e5363cf2d   tada-tts-3b-ml-q8_0.gguf
ef5652e7a346c8a55dd6692676da2827320fd141042e87175880e032e1953082   tada-codec-f16.gguf
7efcc96795dd2b27577a4a81eb52d0c3add5ffa67f325fba5a938f3f98067ace  tada-ref.gguf

Provenance and EU AI Act Art. 53 note

• Upstream model: HumeAI/tada-3b-ml — published by HumeAI.
• Upstream licence: llama3.2. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.