PrimeTTS v2-Stream-Clean — token-level streaming, v2-clean audio
The definitive streaming variant: token-level band-attention encoder (input streams incrementally) + v2's non-causal clean vocoder (no parasite noise). Reconciles token-level input streaming with clean audio — the causal v2-Stream sacrificed quality unnecessarily; this doesn't.
v2streamclean_enc.onnx— text (x,tone,lang,x_lengths,noise_scale,length_scale) → z[1,192,T]. Band encoder, token-level. Run once per phrase.v2streamclean_dec.onnx— z[1,192,Tc] → wav[1,1,Tc·256]. Clean non-causal vocoder. Run per chunk, overlap-save.onnx_stream.py— reference runner (uses the right params).
Streaming params: chunk = 24, left = 64, RIGHT = 16 (the clean non-causal vocoder needs 16 future frames for bit-exact chunking; the causal one used 4). 16 kHz, zh-TW + English.
from onnx_stream import StreamingTTS # RIGHT=16 baked in
tts = StreamingTTS("v2streamclean_enc.onnx", "v2streamclean_dec.onnx")
z = tts.encode(phone_ids, tone_ids, lang_ids) # once
for pcm in tts.stream(z): play(pcm) # per 24-frame chunk, clean audio
Frontend (text→ids): g2pw bopomofo + g2p_en, 88 syms/6 tones/2 langs, add_blank.
sherpa-onnx: OfflineTtsMbistftStreamModel(enc, dec, num_threads=2, right_lookahead=16).
License: Apache-2.0 · part of Luigi/PrimeTTS.