TransWithAI Cohere Timestamp

This app keeps the reference Faster Whisper UI/CLI shape while delegating transcription, VAD, forced alignment, and subtitle writing to a project-built crispasr binary.

The default output is Japanese subtitles: language: "ja" and sub_formats: ["srt"] in generation_config.json5.

Quick Start

scripts/setup.sh cpu
python infer.py --help
python infer.py --dry_run --device cpu path/to/audio_or_folder
python infer.py --device cuda --sub_formats srt,vtt,lrc path/to/audio_or_folder

scripts/setup.sh initializes the CrispASR submodule, compiles the selected backend, downloads the pinned Cohere, Qwen3, and Whisper VAD GGUF models, runs tests, and performs a wrapper dry-run.

Backend choices:

scripts/setup.sh cpu
scripts/setup.sh cuda128
scripts/setup.sh cuda122
scripts/setup.sh cuda118
scripts/setup.sh vulkan

Use --skip-models when you only want to compile CrispASR, or --skip-tests when rebuilding quickly.

Build scripts cap parallel compilation at 16 jobs by default. You can set JOBS=1 through JOBS=16, but higher values are capped to 16.

The wrapper defaults crispasr_path to auto, which checks these local build outputs before falling back to crispasr on PATH:

• runtimes/<platform>/cuda128/bin/crispasr
• runtimes/<platform>/cuda122/bin/crispasr
• runtimes/<platform>/cuda118/bin/crispasr
• runtimes/<platform>/vulkan/bin/crispasr
• runtimes/<platform>/cpu/bin/crispasr
• build/crispasr-cuda128/bin/crispasr
• build/crispasr-cuda122/bin/crispasr
• build/crispasr-cuda118/bin/crispasr
• build/crispasr-vulkan/bin/crispasr
• build/crispasr-cpu/bin/crispasr

Compatibility Flags

Preserved practical flags include --model_name_or_path, --device, --compute_type, --overwrite, --audio_suffixes, --sub_formats, --output_dir, --generation_config, --log_level, VAD override flags, merge segment flags, and positional files/directories.

Configuration precedence is:

1. Built-in defaults.
2. generation_config.json5 values.
3. CLI flags that are explicitly present.

Faster Whisper-only knobs such as --compute_type, fine-grained VAD thresholds, and batching flags are accepted for launcher compatibility but are not reimplemented in Python. CrispASR owns runtime batching, ASR, VAD, alignment, and subtitle writing.

CrispASR Command Shape

For an input file, this wrapper builds commands like:

build/crispasr-cpu/bin/crispasr --backend cohere -m models/cohere/model.gguf -f input.mp3 --vad -vm models/vad/whisper-vad-asmr-q4_k.gguf -am models/qwen3-forced-aligner.gguf --force-aligner -l ja --split-on-punct -of input -osrt

Subtitle format mapping:

• srt -> -osrt
• vtt -> -ovtt
• lrc -> -olrc
• txt -> -otxt
• json -> -oj
• json-full or json_full -> -ojf
• csv -> -ocsv

Device mapping:

• --device cpu -> --gpu-backend cpu
• --device cuda -> --gpu-backend cuda
• --device amd, rocm, or hip -> --gpu-backend vulkan
• --device auto omits --gpu-backend
• gpu_backend in config overrides the automatic mapping

-vm models/vad/whisper-vad-asmr-q4_k.gguf is explicit because CrispASR defaults to Silero VAD and the whisper-vad alias resolves through CrispASR's cache. -am is explicit because CrispASR -am auto maps to canary CTC, not the requested Qwen3 forced aligner.

File Discovery And Outputs

Positional files and directories are accepted. Directories are scanned recursively by suffix. Existing output files are skipped unless --overwrite is provided. When --output_dir is set, relative directory structure is mirrored under that directory; otherwise outputs are written beside the source audio.

Default CrispASR builds support wav, flac, mp3, and ogg. Convert container/video formats such as m4a, mp4, webm, mkv, or mov to one of those audio formats before running this wrapper unless you build CrispASR with additional FFmpeg support yourself.

latest.log is overwritten on every run and captures DEBUG-level logs.

Models

download_models.py reads models/manifest.json5, downloads artifacts over HTTPS, writes only under models/, and requires SHA-256 verification before download. The manifest includes the TransWithAI Japanese Cohere ASR Q6_K model, Qwen3 forced aligner, and Whisper VAD sidecar so release bundles can run without resolving CrispASR's VAD cache alias.

python download_models.py --dry_run
python download_models.py

Submodule

CrispASR is tracked as third_party/CrispASR through .gitmodules. See docs/submodules.md for checkout requirements.

The build scripts compile the executable target crispasr-cli, which outputs a binary named crispasr:

scripts/build-crispasr-cpu.sh
scripts/build-crispasr-cuda118.sh
scripts/build-crispasr-cuda122.sh
scripts/build-crispasr-cuda128.sh
scripts/build-crispasr-vulkan.sh

Vulkan builds require a glslc shader compiler. In Conda environments this is provided by shaderc; on Ubuntu install the glslc package alongside libvulkan-dev and spirv-headers.

CUDA builds are split by runtime line, matching the Faster Whisper release strategy. Each CUDA script is intended to run inside its matching Conda environment with the matching cuda-compiler package, then CI removes compiler packages before staging the runtime artifact:

• environment-cuda118.yml -> scripts/build-crispasr-cuda118.sh
• environment-cuda122.yml -> scripts/build-crispasr-cuda122.sh
• environment-cuda128.yml -> scripts/build-crispasr-cuda128.sh

The wrapper selects the newest compatible CUDA runtime based on the installed NVIDIA driver, then falls back to Vulkan or CPU if no CUDA runtime is usable. Set CUDA_ARCHITECTURES=... only when you intentionally want a custom architecture list for a build.

Release Bundles

The release workflow mirrors the reference Faster Whisper workflow style: build runtime artifacts, package archives, upload CI artifacts, and publish them on v* tags using large-release upload support.

python scripts/package-universal.py --runtime cpu --archive dist/transwithai_cohere_timestamp_linux-x64_cpu.zip
python scripts/package-universal.py --runtime cuda128 --archive dist/transwithai_cohere_timestamp_linux-x64_cuda128.zip
python scripts/package-universal.py --archive dist/transwithai_cohere_timestamp_linux-x64_universal.zip

Full app bundles always include every artifact listed in models/manifest.json5; run python download_models.py before packaging. Unlike the reference Faster Whisper release matrix, this project does not publish an empty-model variant.

Per-runtime bundles are complete downloadable app bundles with the launch .bat files, project source/docs, models/, and exactly one runtime under runtimes/linux-x64/<runtime>/. Release builds publish these runtime-specific archives:

• transwithai_cohere_timestamp_linux-x64_cpu.zip
• transwithai_cohere_timestamp_linux-x64_vulkan.zip
• transwithai_cohere_timestamp_linux-x64_cuda118.zip
• transwithai_cohere_timestamp_linux-x64_cuda122.zip
• transwithai_cohere_timestamp_linux-x64_cuda128.zip

The universal bundle is also published as transwithai_cohere_timestamp_linux-x64_universal.zip and includes all staged runtimes:

runtimes/linux-x64/cpu/bin/crispasr
runtimes/linux-x64/cuda118/bin/crispasr
runtimes/linux-x64/cuda122/bin/crispasr
runtimes/linux-x64/cuda128/bin/crispasr
runtimes/linux-x64/vulkan/bin/crispasr

Each CUDA runtime is built in a separate Conda environment with the appropriate CUDA compiler/runtime line. Pick one per-runtime archive for a smaller direct download, or use the universal archive when you want one bundle that can choose among all available runtimes automatically.

CUDA/Vulkan Notes

This wrapper does not depend on PyTorch, qwen-asr, transformers, or the Cohere SDK. CrispASR is a native binary project, so its support matrix differs from Faster Whisper. See docs/cuda-support.md and the environment sketches for CPU, CUDA 12.8, and Vulkan-first builds.