Noise occurs only when using Irodori-TTS (on HIP; CUDA/Vulkan/CPU not affected)

[tootomocr]

I built and packaged the software following the instructions in docs/build/windows-hip-distribution.md. When I imported Irodori-TTS via the web UI and generated TTS in Japanese, noise began to occur around the 22-second mark. When I tried generating TTS using Pocket TTS instead, the issue did not happen. (I also tried the voice design option, but got the same result)

Environment

• CPU: Ryzen 7 5700X

• System RAM: 32GB

• GPU: Radeon RX 9060 XT (VRAM 16 GB)

• Graphic Driver: AMD Software: PRO Edition 26.Q3

• Chipset Driver: AMD Chipset Drivers 8.08.12.551

Console log:

Starting audio.cpp server (HIP backend)...
Open http://127.0.0.1:8080 in your browser

audio.cpp is optimized for CUDA. The hip server backend is intended for portability and testing, but performance and model coverage may be lower than CUDA.
native WebUI model root: "C:\\AI\\audio.cpp\\models"
native WebUI package resources: "C:\\AI\\audio.cpp"
audiocpp_server listening on http://127.0.0.1:8080
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16304 MiB):
  Device 0: AMD Radeon RX 9060 XT, gfx1200 (0x1200), VMM: no, Wave Size: 32, VRAM: 16304 MiB

.bat file for launch:

@echo off
set PATH=C:\Program Files\AMD\ROCm\7.1\bin;%PATH%
cd /d "%~dp0"
echo Starting audio.cpp server (HIP backend)...
echo Open http://127.0.0.1:8080 in your browser
echo.
audiocpp_server.exe --ui --backend hip --ui-management
pause

Text to speech:

Model	Text	seed
Irodori-TTS 500M v3	皆さん、こんにちは。今日は「挑戦することの大切さ」についてお話しします。新しいことに挑戦するとき、失敗するのではないかと不安になることがあります。しかし、失敗は決して無駄ではありません。挑戦したからこそ得られる経験や学びがあります。小さな一歩でも、勇気を持って踏み出すことが大切です。私もこれから、失敗を恐れず、さまざまなことに挑戦していきたいと思います。ご清聴ありがとうございました。	1234
Irodori-TTS v4 small	same	same
Pocket-TTS	Hello everyone. Today, I would like to talk about the importance of taking on challenges. When we try something new, we may feel worried about making mistakes or failing. However, failure is not always a bad thing. It can teach us important lessons and help us grow. Even a small challenge can give us confidence and new experiences. I believe that taking the first step is more important than being afraid of failure. From now on, I want to try new things, learn from my mistakes, and become a better person. Thank you for listening.	same

Results:

Model	Output	Generation Time	sec
Irodori-TTS 500M v3 Q8	Irodori-TTS-500M-v3-Q8-7.88s.wav	7.88s	29s
Irodori-TTS 500M v3 FP16	Irodori-TTS-500M-v3-FP16-8.65s.wav	8.65s	29s
Irodori-TTS v4 Small Q8	Irodori-TTS-v4-Small-Q8-8.44s.wav	8.44s	28s
Irodori-TTS v4 Small FP16	Irodori-TTS-v4-Small-FP16-8.78s.wav	8.78s	28s
Pocket-TTS Q8	Pocket-TTS-Q8-2.84s.wav	2.84s	27s
Pocket-TTS BF16	Pocket-TTS-BF16-2.84s.wav	2.84s	27s

---

Attachments

• Irodori-TTS-500M-v3-FP16-8.65s.wav --The number of seconds represents the generation time
• Irodori-TTS-500M-v3-Q8-7.88s.wav
• Irodori-TTS-v4-Small-FP16-8.78s.wav
• Irodori-TTS-v4-Small-Q8-8.44s.wav
• Pocket-TTS-BF16-2.84s.wav
• Pocket-TTS-Q8-2.84s.wav

[0xShug0]

@tootomocr This is an interesting bug. On CUDA and Vulkan, I confirmed that your test case produces no noise when run directly via either the CLI or the server. The UI is also not the issue (check the video attached). It looks like some HIP specific behavior.

@IIIIIllllIIIIIlllll Could you take a look when you get a chance?

Update: @fedeizzo I suspect this is related to enabling the HIP ConvTranspose1dModule fast path globally.

simplescreenrecorder-2026-08-18_14.03.14.mp4

irodori_500m_v3_q8_cli.wav
irodori_v4_small_q8_cli.wav
irodori_500m_v3_q8_server.wav
irodori_v4_small_q8_server.wav

irodori_500m_v3_q8_cli_vulkan.wav
irodori_v4_small_q8_cli_vulkan.wav
irodori_500m_v3_q8_server_vulkan.wav
irodori_v4_small_q8_server_vulkan.wav

[tootomocr]

@0xShug0 Thank you for your reply! After that, I tried testing the CPU and Vulkan just to be sure.

Results (server)

CPU: (windows-cpu-balance release-0.6)

Model	Output	configuration
Irodori-TTS v4 Small Q8	irodori_tts_v4_small_q8_output_cpu_server.wav	{"seed":1234,"chunks":1,"characters":191,"chunk_budget":1000,"timings":[{"chunk":1,"characters":191,"wall_ms":"439455.359900","rtf":"14.967826"}]}
Irodori-TTS 500M v3 Q8	irodori_tts_v3_500m_q8_output_cpu_server.wav	I lost it, but I think it's the same one.

Vulkan: (VulkanSDK 1.4.357.0)

Model	Output	configuration
Irodori-TTS v4 Small Q8	irodori_tts_v4_small_q8_output_vulkan_server.wav	{"seed":1234,"chunks":1,"characters":191,"chunk_budget":1000,"timings":[{"chunk":1,"characters":191,"wall_ms":"6902.477100","rtf":"0.241683"}]}
Irodori-TTS 500M v3 Q8	irodori_tts_v3_500m_q8_output_vulkan_server.wav	{"seed":1234,"chunks":1,"characters":193,"chunk_budget":1000,"timings":[{"chunk":1,"characters":193,"wall_ms":"7105.968600","rtf":"0.241371"}]}

No noise was generated when using the CPU and Vulkan. Is there an AMD user out there who could test Irodori-TTS on HIP?

[IIIIIllllIIIIIlllll]

Let me check.

[IIIIIllllIIIIIlllll]

Sorry, could you please try ROCm 6.4:
https://github.com/IIIIIllllIIIIIlllll/audio.cpp/releases/tag/v0.6?
I couldn't reproduce this problem on RDNA3.

[tootomocr]

@IIIIIllllIIIIIlllll Thank you for taking the time out of your busy schedule. I tested both of your builds (7.1 and 6.4), but the same issue still occurred.

[IIIIIllllIIIIIlllll]

@IIIIIllllIIIIIlllll Thank you for taking the time out of your busy schedule. I tested both of your builds (7.1 and 6.4), but the same issue still occurred.

I'm sorry, but since I don't have an RDNA4 GPU, it will be quite difficult to solve, and this could be a long-term task.

[IIIIIllllIIIIIlllll]

Could you give a try? Force RDNA4 using FP16.

set GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F=1
audiocpp_cli.exe --backend hip --task tts --family irodori_tts --model ... 

[tootomocr]

@IIIIIllllIIIIIlllll I gave it a try

---

Command:

set GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F=1
audiocpp_cli.exe --backend hip --task tts --family irodori_tts --model models\Irodori-TTS-v4-Small-GGUF --text "皆さん、こんにちは。今日は「挑戦することの大切さ」についてお話しします。新しいことに挑戦するとき、失敗するのではないかと不安になることがあります。しかし、失敗は決して無駄ではありません。挑戦したからこそ得られる経験や学びがあります。小さな一歩でも、勇気を持って踏み出すことが大切です。私もこれから、失敗を恐れず、さまざまなことに挑戦していきたいと思います。ご清聴ありがとうございました。" --out output\test.wav

Result:

Log:

ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16304 MiB):
  Device 0: AMD Radeon RX 9060 XT, gfx1200 (0x1200), VMM: no, Wave Size: 32, VRAM: 16304 MiB
Detected GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F
family=irodori_tts
task=tts
mode=offline
audio_out=output\test.wav

Output:

test.wav

---

Hmm, that's a difficult problem.
I'll let you know if there are any progress on our side.

[IIIIIllllIIIIIlllll]

Can't reproduce on RDNA3 (780M/gfx1103): same text + seed gives clean 29.4 s audio on both the ROCm 6.4 and 7.1 builds, and GPU-vs-CPU codec output is bit-identical there.

Combined with your results (both ROCm builds affected, GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F no effect, Vulkan fine, Q8+FP16 both affected), the main suspect is now the fattn-tile flash-attention kernel compiled for gfx1200 — it's the one component that's HIP-only, shared by both precisions, compiled per-arch, and position-sensitive (a bad range of query positions would corrupt a time range of the latent, matching the ~22 s onset).

Could you test a build with -DGGML_CUDA_NO_FA=ON (same build_windows_hip.ps1 command as before, just add that flag to the CMake args)? It falls back to decomposed attention — on gfx1103 that path is bit-identical to the fused kernel, so it's a clean A/B test. Noise gone → fattn-tile on gfx1200 confirmed; still there → next suspects are the RDNA4 WMMA matmul paths.

Also useful: does a short text (~10–15 s of audio) come out clean? That tells us whether the noise starts at an absolute ~22 s or at a fixed fraction of the output.

[tootomocr]

@IIIIIllllIIIIIlllll Okay

Changes:

build_windows_hip.ps1

$configureArgs = @(
    "-S", $sourceDir,
    "-B", $buildDir,
    "-G", "Ninja",
    "-DCMAKE_MAKE_PROGRAM=$(Convert-ToCMakePath $ninja)",
    "-DCMAKE_BUILD_TYPE=$BuildType",
    "-DCMAKE_C_COMPILER=$(Convert-ToCMakePath $clangc)",
    "-DCMAKE_CXX_COMPILER=$(Convert-ToCMakePath $clangxx)",
    # GPU_BUILD_TARGETS/AMDGPU_TARGETS are sticky cache entries set by hip-config-amd.cmake
    # from GPU_TARGETS; without -U a previous configure's arch list wins over a new -GpuTargets.
    "-UGPU_BUILD_TARGETS",
    "-UAMDGPU_TARGETS",
    "-DENGINE_ENABLE_CUDA=OFF",
    "-DENGINE_ENABLE_HIP=ON",
    "-DENGINE_ENABLE_OPENMP=OFF",
    # ggml probes OpenMP on its own (GGML_OPENMP defaults ON) and would silently
    # pick up libomp from a vcvars environment, adding a libomp140.x86_64.dll
    # runtime dependency; force it off so the dependency list stays minimal.
    "-DGGML_OPENMP=OFF",
    "-DGPU_TARGETS=$GpuTargets",
    "-DGGML_HIP_HIPBLASLT=$hipblasLtValue",
    "-DGGML_CUDA_FORCE_MMQ=$forceMmqValue",
    "-DGGML_HIP_NO_VMM=$noVmmValue",
    "-DENGINE_ENABLE_NATIVE_CPU=$nativeCpuValue",
    "-DAUDIOCPP_DEPLOYMENT_BUILD=$deploymentValue",
    "-DENGINE_ENABLE_CUDA_GRAPHS=$graphsValue",
HERE--> "-DGGML_CUDA_NO_FA=ON"
)

---

I tested it using Irodori-TTS-v4-Small-GGUF

Short text (16s)

Command:

audiocpp_cli.exe --backend hip --task tts --family irodori_tts --model models\Irodori-TTS-v4-Small-GGUF --text "皆さん、こんにちは。今日は「挑戦することの大切さ」についてお話しします。新しいことに挑戦するとき、失敗するのではないかと不安になることがあります。しかし、失敗は決して無駄ではありません。" --out output\test.wav

Output:

test_short.wav

Long text (28s)

Command:

audiocpp_cli.exe --backend hip --task tts --family irodori_tts --model models\Irodori-TTS-v4-Small-GGUF --text "皆さん、こんにちは。今日は「挑戦することの大切さ」についてお話しします。新しいことに挑戦するとき、失敗するのではないかと不安になることがあります。しかし、失敗は決して無駄ではありません。挑戦したからこそ得られる経験や学びがあります。小さな一歩でも、勇気を持って踏み出すことが大切です。私もこれから、失敗を恐れず、さまざまなことに挑戦していきたいと思います。ご清聴ありがとうございました。" --out output\test.wav

Output:

test_long.wav

---

We found that the same problem occurs, but not in short texts.