# Research-list qualification matrix

The matched local pack is four clips totaling 444.192 s. Times include one model load per batch unless noted. A blocked or unsupported runtime is recorded rather than silently substituted.

| Lane | Exact implementation/model | ROCm qualification | CPU qualification |
|---|---|---|---|
| whisper.cpp | upstream `ggml-org/whisper.cpp`, `large-v3-turbo` GGUF | 11.027 s; upstream HIP build on `gfx1201` | 251.119 s with GPU offload disabled |
| faster-whisper | CTranslate2 4.8.2 source build with `WITH_HIP`, `large-v3-turbo` | 33.034 s, float32. `CT2_CUDA_ALLOCATOR=cub_caching` is required: the default allocator reproduced a gfx1201 GPU page fault. | 150.357 s, CTranslate2 int8 |
| WhisperX | WhisperX 3.8.6, `large-v3-turbo`, full alignment | 33.412 s, float32 ASR plus PyTorch ROCm alignment. Uses the same CTranslate2 HIP build and CUB allocator workaround. | 106.740 s, int8 ASR plus alignment |
| Parakeet Redux | `moondream/parakeet-redux` via Photon | Unsupported: installed GPU runtime requires NVIDIA CUDA | 11.790 s |
| Parakeet Ultra | `moondream/parakeet-ultra` via Photon | Unsupported: installed GPU runtime requires NVIDIA CUDA | 37.381 s |
| Qwen3 ASR | `Qwen/Qwen3-ASR-1.7B` | 46.899 s, PyTorch ROCm 7.2 | 350.263 s |
| Qwen3 ASR + ForcedAligner | Qwen3 ASR plus `Qwen/Qwen3-ForcedAligner-0.6B` | 75.125 s, transcript and word timestamps; 4.57 GiB PyTorch peak | 459.267 s, transcript and word timestamps |
| VibeVoice Streaming 7B | `microsoft/VibeVoice-ASR-Streaming-7B` | Blocked: model load exhausted the 15.92 GiB device; official release has no quantized checkpoint | Blocked: matched pack did not complete in one hour |
| VibeVoice long-form | `microsoft/VibeVoice-ASR` | Not downloaded: a bounded-state driver is required before testing; the known full-input streaming wrapper is unsuitable for the 60-minute recording | Same blocker |
| VibeVoice BitNet | `microsoft/VibeVoice-ASR-BitNet` / VibeASR.cpp | CPU-only official runtime | 116.876 s |
| Insanely Fast Whisper | `openai/whisper-large-v3` through `insanely-fast-whisper` | Unsupported: upstream NVIDIA/MPS-only | Unsupported: upstream has no CPU mode |
| Sortformer reference | `nvidia/diar_sortformer_4spk-v2.1` | Blocked: model access is gated; no terms were accepted | Same blocker |
| ECAPA baseline lane | `speechbrain/spkrec-ecapa-voxceleb` | 11.434 s for 149 three-second embeddings; official matching PyTorch/torchaudio ROCm 7.2 wheels | 18.584 s, same 149 windows |

## ROCm repair evidence

CTranslate2's documented source build supports `WITH_HIP`; the local build targets `gfx1201` and uses ROCm 7.2. The default CTranslate2 allocator reproduced the exact `Memory access fault by GPU node-1` failure tracked upstream for RX 9070 XT/gfx1201 in [OpenNMT/CTranslate2#2021](https://github.com/OpenNMT/CTranslate2/issues/2021). That issue's confirmed workaround, `CT2_CUDA_ALLOCATOR=cub_caching`, completed both faster-whisper and WhisperX matched-pack runs here. The runner scripts reject the unsafe default configuration.

SpeechBrain itself only requires Torch and torchaudio. The prior failure was resolver selection of a CUDA torchaudio wheel, not an ECAPA limitation. Installing matching `2.11.0+rocm7.2` Torch and torchaudio wheels produced a real ROCm tensor smoke test and the reported ECAPA benchmark.