Qwen3-TTS PR 5: 1.7B CustomVoice, VoiceDesign and Base (voice clone) - #302

Open

merceod wants to merge 110 commits into

mainfrom

model/qwen3-tts

Commits on Sep 22, 2026

sdk: reference audio for client.tts voice cloning
merceod
committed
5ac3be0
test: client.tts forwards voice and reference audio
merceod
committed
48c06d9
qwen3_tts: variant-aware config with speaker-encoder settings
merceod
committed
a796d26
qwen3_tts: project Talker-width inputs into the code predictor
merceod
committed
28a34d9
qwen3_tts: build prefill with instruct, optional speaker and both text layouts
merceod
committed
3259e93
qwen3_tts: serve CustomVoice, VoiceDesign and Base from one class
merceod
committed
cbb27df
Register the Qwen3-TTS 1.7B variants
merceod
committed
83c160b
cli: default configs and hints for the Qwen3-TTS 1.7B variants
merceod
committed
5092cd4
Add the Qwen3-TTS 1.7B CustomVoice deployment config
merceod
committed
d63c7c9
Add the Qwen3-TTS VoiceDesign deployment config
merceod
committed
297ebed
Add the Qwen3-TTS Base deployment config
merceod
committed
db2af3d
api: OpenAI speech adapter for Qwen3-TTS
merceod
committed
13022b7
benchmark: Qwen3-TTS 1.7B and VoiceDesign model entries
merceod
committed
f6410a3
benchmark: route Qwen3-TTS through /v1/audio/speech
merceod
committed
3a87ece
test: Qwen3-TTS variants, prefill layouts and predictor projection
merceod
committed
8f32c65
test: Qwen3-TTS speech adapter
merceod
committed
20db9ee
test: Qwen3-TTS 1.7B real-weight loading checks
merceod
committed
1b66b1e
benchmark: streaming /v1/audio/speech client shared by all TTS engines
merceod
committed
6ba7e92
benchmark: WER scorer for synthesized speech
merceod
committed
b4c4fa6
test: Qwen3-TTS parity harness against the qwen-tts reference
merceod
committed
ce08c3d
qwen3_tts: ECAPA-TDNN speaker encoder and mel front end components
merceod
committed
02799bf
test: Qwen3-TTS speaker encoder parity against qwen-tts
merceod
committed
12e5f87
qwen3_tts: codec encoder settings and reference frame arithmetic
merceod
committed
124fb13
qwen3_tts: reference encoder node, in-context clone prefill, codec trimming
merceod
committed
b708a57
qwen3_tts: Base voice-clone walks and CPU-built codec modules
merceod
committed
6439f21
api: reference audio for Qwen3-TTS speech requests
merceod
committed
9b3453e
configs: map the Qwen3-TTS RefEncoder node
merceod
committed
6d3e80b
test: Qwen3-TTS voice-clone prompt, prefill and codec trimming
merceod
committed
29722cc
test: Qwen3-TTS reference audio through the speech adapter
merceod
committed
7cc5872
test: float32 reference codec in the Qwen3-TTS parity harness
merceod
committed
c0bdd17
docs: Qwen3-TTS 1.7B variants, voice cloning and benchmark client
merceod
committed
82cc576
examples: Qwen3-TTS SDK usage for the three variants
merceod
committed
f327647
test: Qwen3-TTS rejects reference audio on non-Base checkpoints
merceod
committed
6e65cb5
streaming: scheduled left-context chunk policy for early first audio
merceod
committed
f60d139
streaming: report each chunk's already-delivered context items
merceod
committed
c5a123e
graph: carry stream chunk offset and context on synthetic edges
merceod
committed
cd61dfe
worker: expose stream chunk geometry to consumers via step_metadata
merceod
committed
3534f50
test: scheduled chunk policy and chunk context geometry
merceod
committed
796651b
qwen3_tts: ramped codec chunk schedule with 25-frame left context
merceod
committed
e1b8584
qwen3_tts: stream codec frames through the scheduled chunk policy
merceod
committed
f4940f7
qwen3_tts: codec windows padded to captured buckets, context from the stream
merceod
committed
616ab0f
test: Qwen3-TTS codec ramp, bucket padding and context trimming
merceod
committed
9b36195
docs: Qwen3-TTS codec chunk ramp and capture buckets
merceod
committed
f27ce58
qwen3_tts: reference encoder keeps its float32 front ends under autocast
merceod
committed
ae11671
test: voice-clone mode in the Qwen3-TTS parity harness
merceod
committed
7c22796
qwen3_tts: advertise the codec clone walk only on Base checkpoints
merceod
committed
29691b5
test: codec graph configs replay the clone walk only for Base
merceod
committed
81c0514
qwen3_tts: pick the codec capture bucket from stream metadata
merceod
committed
52b047e
test: codec capture key from stream metadata
merceod
committed
f4ac432
streaming: report each chunk's item count
merceod
committed
22013c4
graph: carry the stream chunk item count on synthetic edges
merceod
committed
f7d58e8
worker: include the chunk item count in stream_chunks metadata
merceod
committed
3972df8
test: stream chunk item counts
merceod
committed
3818263
deps: soundfile in the qwen3_tts extra for reference audio
merceod
committed
c1ef190
qwen3_tts: hand the codec its reference frame count from the first clone chunk
merceod
committed
36efac1
test: codec clone walk receives ref_frames on its initial inputs
merceod
committed
bc5b6d3
api: sentence splitter for long text-to-speech inputs
merceod
committed
752611e
api: per-adapter sentence-chunking thresholds for /v1/audio/speech
merceod
committed
e234e94
api: synthesize long speech inputs as ordered sentence chunks
merceod
committed
afc57df
test: sentence chunking on /v1/audio/speech
merceod
committed
0b77838
api: sentence-chunk Qwen3-TTS speech inputs from 600 characters
merceod
committed
bf73d67
test: Qwen3-TTS speech chunking thresholds
merceod
committed
dacf98b
docs: Qwen3-TTS sentence chunking on /v1/audio/speech
merceod
committed
46562e6
test: parity harness applies the deployment's resource overrides
merceod
committed
bc4aa0f
benchmark: render the TTS protocol tables from result JSONs
merceod
committed
d3b8e60
test: end-to-end smoke test for a served Qwen3-TTS variant
merceod
committed
9a8d8e2
test: parity harness tolerates specs without a config
merceod
committed
5b27bcd
test: 1.7B real-weight layouts for the five-field prompt layout and a clone prefill check
merceod
committed
f5ac4e2
qwen3_tts: load codec quantizer buffers and check coverage against the state dict
merceod
committed
5396fdb
test: VoiceDesign real-weight layout check with an explicit language
merceod
committed
316ac5a
test: parity harness keeps whole frames in the greedy loop
merceod
committed
3a9c156
test: parity verdict on confident-position agreement and relative logit error
merceod
committed
33ada0c
benchmark: report confident agreement and relative logit error in the parity table
merceod
committed
e7866d5
test: clone parity compares codec encoders in both dtypes and shares reference codes
merceod
committed
61d6bfa
api: streaming speech surfaces an up-front engine error as its HTTP status
merceod
committed
0c4d7c0
test: streaming speech returns the engine error status
merceod
committed
edd23dc
test: parity reports token diversity and loudness of each side's greedy run
merceod
committed
b1e2913
qwen3_tts: memoise reference-clip conditioning by content
merceod
committed
60e5961
test: reference-clip conditioning memo
merceod
committed
9befd9b
docs: reference clips are encoded once per distinct clip
merceod
committed
7dc9746
qwen3_tts: codec window geometry travels with the pass, not request state
merceod
committed
3171700
test: codec trims each pass with its own window geometry
merceod
committed
9262ccb
benchmark: order report rows by system and concurrency, trim the environment dump
merceod
committed
5affa75
qwen3_tts: codec batches mixed windows and returns per-request samples from a single forward
merceod
committed
22de57d
test: codec mixed-window batching and single-forward output shape
merceod
committed
10b293b
qwen3_tts: codec batches up to 32 windows
merceod
committed
0ba9b8f
test: codec batch limit of 32
merceod
committed
36afc51
docs: codec graphs captured for batch sizes up to 32
merceod
committed
fc54067
qwen3_tts: first codec window is a single frame
merceod
committed
f55e16e
qwen3_tts: graph docstring follows the 1-3-8-16 chunk ramp
merceod
committed
7b3381e
qwen3_tts: submodule header follows the 1-3-8-16 chunk ramp
merceod
committed
f342ff5
docs: first audio after a single codec frame
merceod
committed
e8b7fca
test: smoke repeats a seeded sampled run for Base, greedy for the others
merceod
committed
b4fff0b
benchmark: WER loads audio with soundfile, no ffmpeg needed
merceod
committed
45c7b15
conductor: optional rank_devices mapping places worker ranks on a shared device
merceod
committed
9042454
docs: rank_devices for several workers on one GPU
merceod
committed
03cd277
test: worker device mapping
merceod
committed
7b4a979
qwen3_tts: deployment with the codec on a second worker sharing the GPU
merceod
committed
16f3320
benchmark: WER drives Whisper directly from 16 kHz arrays
merceod
committed
a44e180
qwen3_tts: talker prefill replays a packed CUDA graph
merceod
committed
35ed040
test: talker prefill graph capture config
merceod
committed
27ebc0d
docs: talker prefill is graph-captured
merceod
committed
88479dd
qwen3_tts: codec_dtype deployment option (float32 default, bfloat16 opt-in)
merceod
committed
58eca9a
qwen3_tts: codec batch comment matches the 32-request ceiling
merceod
committed
44cdb5c
docs: codec_dtype option
merceod
committed
fe6d9ad
test: codec_dtype parsing
merceod
committed
ac890ef
qwen3_tts: clone stream carries only the reference tail the codec can use as context
merceod
committed
5fca373
test: clone stream reference tail and trimming
merceod
committed
8c90721
qwen3_tts: split deployment caps the KV pool for two workers on one GPU
merceod
committed
2a8fa50
docs: split deployment and CUDA MPS note
merceod
committed
708f72d