TriStream-SVS is a 321.8M-parameter singing-voice synthesis model built from scratch at CorX Labs. It encodes a performance in three streams that mirror the source-filter theory of voice production: a source stream that sees only pitch, a filter stream that alone carries timbre and singer identity, and a residual stream for breath and fricatives. They meet only in a late fusion trunk whose representation is pushed, through a gradient-reversal singer classifier, to forget who is singing. A contrastive speaker encoder enables zero-shot cloning, and a rectified-flow decoder generates 24 kHz mel spectrograms in about 32 steps. Editing masked spans is the native training task and generation is its fully masked case. Validation is singer-disjoint throughout. No listening tests or objective scores have been run, and none are claimed.
The source-filter theory of speech [1] factors a voice into three parts: a glottal source, the vibration of the vocal folds that becomes perceived pitch; a vocal-tract filter, the resonances of throat, mouth and tongue that make a voice sound like a particular person and a vowel sound like that vowel; and turbulent noise, the breath and fricatives. Most singing synthesisers leave a network to discover this structure, or encourage it with an extra loss. TriStream-SVS builds it into the wiring: three encoder streams that cannot exchange information until a late fusion trunk. This report describes the architecture, the training methodology and the current status. It reports no listening tests and no objective scores, because none have been run.
All transformer blocks share one shape from the released config.json: width 768, 12 heads, feed-forward
width 3,072, rotary positions [2], SwiGLU [3] and RMSNorm [4].
| Component | Layers | Receives | Models |
|---|---|---|---|
| Source stream | 3 | log F0 contour, voiced/unvoiced flag | the glottal pulse train: pitch, nothing else |
| Filter stream | 5 | mel spectrogram, singer and emotion/technique embeddings, lyric cross-attention | formants, timbre and identity |
| Residual stream | 2 | mel spectrogram | breath noise, fricatives, articulation texture |
| Lyric encoder | 4 | phonemes (512-symbol inventory) | what is sung, kept apart from how it sounds |
| Fusion trunk | 8 | the three streams, lyric cross-attention | the one place they interact |
| Decoder | 5 | fused representation | rectified-flow velocity field to a 100-band mel |
Disentanglement by construction. The source stream never receives the spectrogram or any speaker embedding, so there is no path along which identity could reach it, whatever the training pressure. The filter stream is the only one that carries identity, which makes "who the voice sounds like" an isolated component that can be inspected.
Speaker encoder. A separate convolutional network turns an arbitrary reference clip into a fixed-size voice embedding using dilated convolutions, squeeze-excitation blocks and attentive statistics pooling, in the spirit of ECAPA-TDNN [5]. It is trained contrastively, so that clips of the same speaker sit close together; this is what allows zero-shot cloning of voices outside the training set.
Figure 1. TriStream-SVS. Inputs on the left feed three streams that meet only in the fusion trunk. Dashed lines are cross-attention and the reversed gradient of the adversarial head.
Adversarial singer head. A classifier over the training singers (a 4,096-entry table) reads the trunk's output through a gradient-reversal layer [6]. As it gets better at naming the singer, the trunk is pushed to make that harder, so the shared representation encodes vocal mechanics rather than individual fingerprints. Here domain-adversarial training is used against memorisation rather than for domain adaptation.
Duration, pitch and beat. Durations are predicted per phoneme in log space, because a sustained vowel can last dozens of times longer than a consonant. Pitch is predicted as separate parts (base note, the scoop into it, and vibrato depth and rate that ramp in over about a quarter second), so each can be adjusted afterwards. With a backing track, a beat tracker pulls durations toward the downbeats by soft quantisation with an adjustable swing, because perfectly quantised timing is one of the clearest tells of synthetic vocals.
Decoder. A five-layer rectified-flow decoder [7, 8] predicts a velocity field from noise to the target mel spectrogram and generates in about 32 steps, against the hundred or more denoising steps of diffusion singers such as DiffSinger [9].
Editing is the native training task: random contiguous spans of real audio are masked and reconstructed from the surrounding audio and the lyrics. Generation reuses every component, starting from a fully masked target: the speaker encoder supplies the voice, the duration and pitch predictors invent timing and melody, and the decoder fills the whole span. "Regenerate this phrase with new lyrics" is therefore not a special case; it is what the network was trained to do.
TriStream-SVS can reproduce a person's vocal characteristics from a short clip, including people it was never trained on. Clone a voice only with the consent of the person it belongs to; do not use it to impersonate anyone or to pass generated singing off as a real artist's. The adversarial head limits memorisation of training singers; it does nothing about misuse of the cloning path at inference time.
Weights and config are released under Apache-2.0 at huggingface.co/CorxLabs/TriStream-SVS-300M.