Views
No views yet
Note: This model is at 18,000 / 100,000 training steps. Quality improves with continued training.
| Property | Value |
|---|---|
| Language | Urdu (اردو) |
| Task | Text-to-Speech |
| Architecture | Rectified Flow DiT |
| Text Backbone | Qwen/Qwen3.5-0.8B |
| Codec | Semantic DACVAE (32-dim) |
| Model Dimension | 1280 |
| Layers | 12 |
| Parameters | ~400M |
| Precision | BF16 |
| Training Steps | 18,000 |
| Validation Loss | 0.7959 |
pip install torch safetensors transformers huggingface_hub soundfile dacvae1from irodori_tts.inference_runtime import InferenceRuntime, RuntimeKey, SamplingRequest
2
3key = RuntimeKey(
4 checkpoint="mahwizzzz/deepspeak-v1",
5 model_device="cuda",
6 model_precision="bf16",
7 codec_device="cuda",
8)
9
10runtime = InferenceRuntime.from_key(key)
11
12result = runtime.synthesize(
13 SamplingRequest(
14 text="یہ ایک آزمائشی جملہ ہے۔",
15 ref_wav="reference.wav",
16 seconds=10.0,
17 num_steps=40,
18 cfg_scale_text=3.0,
19 cfg_scale_speaker=5.0,
20 )
21)
22
23import soundfile as sf
24sf.write("output.wav", result.audio.squeeze(0).numpy(), result.sample_rate)1python infer.py \
2 --checkpoint mahwizzzz/deepspeak-v1 \
3 --text "یہ ایک آزمائشی جملہ ہے۔" \
4 --ref-wav reference.wav \
5 --output-wav output.wav \
6 --model-device cuda \
7 --model-precision bf16
| Parameter | Value |
|---|---|
| Optimizer | Muon |
| Learning Rate | 1e-4 |
| LR Scheduler | WSD (Warmup-Stable-Decay) |
| Warmup Steps | 2,000 |
| Effective Batch Size | 64 |
| Precision | BF16 |
| Max Sequence Length | 256 tokens |
| Latent Steps | 750 |
1@misc{deepspeak2026,
2 title = {DeepSpeak-v1: Urdu Text-to-Speech with Rectified Flow DiT},
3 author = {mahwiz Khalil},
4 year = {2026},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/mahwizzzz/deepspeak-v1}
7}