Cited, not measured on the local desktop
Local speech means a file on your box turns audio into text, or text into audio, without a hosted API. On AMDai the first STT engine is whisper.cpp with Vulkan — the same ggml family as llama.cpp, not a chat /v1. Piper is the first TTS engine (pip install piper-tts). llama.cpp stays first for local LLMs. Speeds stay not yet measured by AMDai.
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
If the app is not installed, packages are not published yet. There is no public download URL. Read the Stack Lab install notes . If Stack Lab is already running on this machine, open the local app .
STT: you feed a wav to whisper.cpp and get text. TTS: you feed a sentence to Piper and get a wav. That is a different job from a local Qwen chat. whisper.cpp files are the ggml models the README lists — not house 35B Q4–Q6 GGUFs. Q3 stays a labeled estimate on the LLM side. Nothing below Q3. We have no speech run card yet.
Disk and mem cells copy the whisper.cpp README Memory usage table. They are not AMDai measurements and not tok/s.
| Model | Disk (cited) | Mem (cited) | Note |
|---|---|---|---|
| tiny | 75 MiB | ~273 MB | Official whisper.cpp README Memory usage table. Smallest listed file. |
| base | 142 MiB | ~388 MB | Official whisper.cpp README Memory usage table. |
| small | 466 MiB | ~852 MB | Official whisper.cpp README Memory usage table. |
| medium | 1.5 GiB | ~2.1 GB | Official whisper.cpp README Memory usage table. |
| large | 2.9 GiB | ~3.9 GB | Official whisper.cpp README Memory usage table. Largest listed file. |
Classes use AMD-published VRAM stickers from the GPU directory. Every fit cell is an estimate against the official mem table. Every speed cell is not yet measured by AMDai.
| VRAM class | Directory pages | Fit (estimate) | Speed |
|---|---|---|---|
| 32 GB discrete | Estimate. The official large row is ~3.9 GB mem — well under a 32 GB sticker. Runtime and quality stay unmeasured here. | not yet measured by AMDai | |
| 24 GB discrete | Estimate. Official table rows through large (~3.9 GB mem) sit under 24 GB. Confirm the sticker. Speed stays unmeasured. | not yet measured by AMDai | |
| 20 GB discrete | Estimate. Official table through large is still under 20 GB. Confirm the sticker before you assume headroom next to a local LLM. | not yet measured by AMDai | |
| 16 GB discrete | Estimate. Official table through large (~3.9 GB mem) sits under 16 GB. Sharing the card with a 35B GGUF is a calculator question, not a speech-mem claim. | not yet measured by AMDai | |
| 12 GB discrete | Estimate. Official large ~3.9 GB mem is under 12 GB. Confirm the sticker. Do not stack a house 35B Q4–Q6 file on the same VRAM without the calculator. | not yet measured by AMDai | |
| 8 GB discrete | Estimate. Official large ~3.9 GB mem is under 8 GB on paper. Start smaller (tiny / base / small) if the card also holds an OS desktop. This is not the 16 GB RX 7600 XT. | not yet measured by AMDai | |
| Shared memory (Strix Halo) | Estimate. This APU drinks from the same DIMMs the OS uses. Official whisper.cpp mem rows are small next to a 64 / 128 GB kit — still not a measured AMDai card. Size the kit on the Strix Halo sheet. | not yet measured by AMDai |
1 · Cited, not measured on the local desktop
Bare-metal Ubuntu 24.04+ is the AMDai measured path for ROCm on the local desktop. Speech follows that OS even though we have no first-party speech card. See the GPU in rocminfo before you pick a HIP gfx pin. Windows is a cited second path, not the first try.
2 · Cited, not measured on the local desktop
whisper.cpp is the AMDai first STT path. The official README documents Vulkan as the cross-vendor GPU build: cmake -B build -DGGML_VULKAN=1 then cmake --build build -j --config Release. That is the same ggml Vulkan idea as llama.cpp — it is not llama-server /v1, and it is not a chat engine. Download a listed ggml model from the project. Real-time factor stays not yet measured by AMDai.
3 · Cited, not measured on the local desktop
The same README documents AMD ROCm/HIP: cmake -B build -DGGML_HIP=1 -DAMDGPU_TARGETS="gfx1201" then cmake --build build -j --config Release. Replace gfx1201 with the string rocminfo prints. The README names gfx1100 (RX 7900 XTX), gfx1101 (RX 7800 XT), and gfx1201 (RX 9070 XT) as common examples — the house R9700 is also gfx1201. Do not mix a gfx110x pin onto a gfx1201 card. HIP is the second STT path after Vulkan.
4 · Cited, not measured on the local desktop
Piper is the AMDai first TTS path. The OHF-Voice/piper1-gpl README documents pip install piper-tts. Voices and the CLI live in that repo. The Python API documents use_cuda=True with onnxruntime-gpu — that is NVIDIA CUDA, not a ROCm wheel AMDai will invent. On Radeon, treat Piper as the documented CPU/ONNX install unless a cited ROCm path appears. Speeds stay not yet measured by AMDai.
Only steps whisper.cpp or Piper document. Everything else is marked unofficial.
Vulkan is the cross-vendor GPU build the whisper.cpp README documents. That is the honest Windows STT path AMDai will name — same idea as llama.cpp Vulkan on /setup/windows. HIP/ROCm userspace stays a Linux product. Cited, not measured on the local desktop. We do not invent a Windows real-time factor.
pip install piper-tts is what the Piper README documents. Do not invent a Windows ROCm Piper pin. Ubuntu 24.04+ on bare metal is still the OS that matches local desktop. WSL2 ROCm is a technical preview on the 10.1 notes — not a measured speech path.
Community wrappers, DirectML forks, and translation layers that whisper.cpp or Piper do not document stay unofficial. AMDai does not walk them and does not invent a Windows real-time factor. If the official README does not list the step, stop.
A local LLM on llama.cpp Vulkan and whisper.cpp Vulkan can share a box. They do not share a /v1. On one card, VRAM is one pool — do not invent a split. Size the LLM in the calculator; whisper.cpp memory stays the official README table. Minimum system RAM needed and RAM needed at full context are the shopper RAM lines for the LLM. A 64 GB kit estimate is the consumer shopper default (8 GB OS/desktop plus 16 GB headroom plus whatever the LLM offload needs) — labeled estimate, not a measured speech peak. Piper is a separate CPU/ONNX process unless a cited ROCm path appears. On a second card, read /multi-gpu. Stack Lab’s Speech (STT + TTS) workload is the companion — Lab Notes hands off with amdai-stack-lab:speech, not a house 35B recipe id. We have not measured speech-plus-LLM on the local desktop.
Shopper RAM labels stay Minimum system RAM needed and RAM needed at full context for the LLM side. Open the calculator for that load. A second card is a /multi-gpu question, not an invented VRAM split.
whisper.cpp Vulkan needs a working Vulkan ICD. If the binary never sees the GPU, you are on CPU. This is the same class of miss as llama.cpp Vulkan — not a ComfyUI HIP/ROCm PyTorch miss. Check the driver and vulkaninfo before you rebuild.
A HIP build that omits the current gfx fails the same way llama.cpp HIP does. The whisper.cpp README says replace gfx1201 with rocminfo | grep gfx. HSA_OVERRIDE_GFX_VERSION is unofficial. Current gfx1201 cards should not need it on a matching build.
Official whisper.cpp large mem is ~3.9 GB. That can still starve a card that already holds a 35B GGUF. Size the LLM on the calculator first. Do not invent a split. A second card is a /multi-gpu question.