Qwen3 TTS Developer Pack Overview#
Included runtime#
The Qwen3 TTS pack contains:
Signed Production and functional Release Qwen3 TTS plugins for CUDA, Vulkan, and D3D12.
nvigi.tts.qwen.execommand-line sample andnvigi.3d.exe3D sample.Qwen3 0.6B Base Talker assets (Q4_K_M default and optional Q8_0) and the shared tokenizer/codec (F32) GGUF asset.
An opt-in provisioning script that downloads Qwen’s hosted example WAV and creates a reusable voice JSON locally; neither artifact is bundled.
Qwen-specific source, public headers, symbols, and HTML documentation.
NVIGI Core/HWI is installed by setup_sample.bat from the public NVIGI Core
1.6.1 runtime. It is not duplicated in the initial package payload.
Model layout#
Item |
Location |
Purpose |
|---|---|---|
Q4_K_M model |
|
Q4_K_M Talker, F32 codec, and model metadata |
Q8_0 model |
|
Q8_0 Talker, F32 codec, and model metadata |
Q4_K_M is the default model GUID. Q8_0 is selected with its distinct GUID; it uses more disk/VRAM and can be slower. Each GUID directory is a complete NVIGI model containing one Talker, the shared F32 codec payload, and independent discovery metadata.
Runtime API#
The public API is in include\nvigi_tts_qwen3.h. A host selects one of the
CUDA, Vulkan, or D3D12 Qwen plugin IDs, creates an inference instance with the
model root, and submits a Qwen prompt data slot.
For voice cloning, provide either:
A precomputed JSON decoded into
refSpkEmband optional ICLrefCodes, orA source WAV through
refAudioPath.
The programming guide describes the exact structures and lifetime rules.
Samples and source#
Sample |
Binary |
Source |
|---|---|---|
CLI |
|
|
3D UI |
|
|
The 3D sample supports Qwen TTS directly. Full ASR -> GPT -> TTS applications are integration scenarios using a separately provisioned NVIGI Developer SDK.
Seed-TTS English quality and performance#
The following baselines were measured with an NVIDIA GeForce RTX 4090 and driver 610.88. It uses the English Seed-TTS evaluation set, with each request conditioned on its corresponding reference WAV and reference transcript through Qwen Base Mode B voice cloning.
Each run used the shipped Q4_K_M v0 Talker and F32 codec, random sampling
(seed = -1), Talker and Code Predictor temperature 0.9, a 768-frame
generation limit, and a 40-word chunk limit. The model was loaded once per
backend and reused for the full batch.
Quality#
Metric |
Result |
|---|---|
English macro WER |
2.008% |
Mean speaker similarity (SIM) |
0.6748 |
Median speaker similarity (SIM) |
0.6825 |
Metric |
Result |
|---|---|
English macro WER |
2.094% |
Mean speaker similarity (SIM) |
0.6742 |
Median speaker similarity (SIM) |
0.6813 |
Metric |
Result |
|---|---|
English macro WER |
2.099% |
Mean speaker similarity (SIM) |
0.6718 |
Median speaker similarity (SIM) |
0.6804 |
WER uses Whisper Large V3 with the official Seed-TTS English normalization and unweighted per-utterance averaging. SIM is the cosine similarity between the official WavLM-large/ECAPA speaker embeddings of each generated WAV and its reference WAV.
Performance#
Measurement |
Mean |
Median |
P95 |
|---|---|---|---|
Time to first codec frame (TTFA) |
164.0 ms |
158.8 ms |
226.0 ms |
Engine-reported full inference |
465.4 ms |
457.6 ms |
635.2 ms |
Real-time factor (RTF) |
0.112 |
0.109 |
0.141 |
Aggregate throughput was 9.16x real time. The complete Seed-TTS generation pass, including one-time model/framework startup and all requests, took 533.6 seconds.
Measurement |
Mean |
Median |
P95 |
|---|---|---|---|
Time to first codec frame (TTFA) |
253.4 ms |
238.7 ms |
370.3 ms |
Engine-reported full inference |
812.7 ms |
797.8 ms |
1,117.2 ms |
Real-time factor (RTF) |
0.197 |
0.191 |
0.248 |
Aggregate throughput was 5.21x real time. The complete Seed-TTS generation pass, including one-time model/framework startup and all requests, took 948.2 seconds.
Measurement |
Mean |
Median |
P95 |
|---|---|---|---|
Time to first codec frame (TTFA) |
180.6 ms |
177.0 ms |
247.9 ms |
Engine-reported full inference |
570.8 ms |
561.3 ms |
779.7 ms |
Real-time factor (RTF) |
0.137 |
0.134 |
0.170 |
Aggregate throughput was 7.47x real time. The complete Seed-TTS generation pass, including one-time model/framework startup and all requests, took 646.6 seconds.
TTFA is measured to the first generated codec-frame codes, not to the first streamed PCM callback. Per-request inference measurements exclude one-time model loading.
Because seed = -1 selects a new random seed, this is a measured single-run
baseline rather than a deterministic performance or quality guarantee. These
figures are NVIDIA-measured baselines; the evaluator is not included in this
package. The WER and speaker-similarity methodology follows ByteDance’s
Seed-TTS evaluator at revision 752f4297f090c46bb1a55a1f7439e5944ddefe8d.
Results may vary with hardware, software, model, and sampling conditions.