0.5.5 - 2026-06-08

Release Notes

Install whisperforge 0.5.5

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/bevsxyz/WhisperForge/releases/download/v0.5.5/whisperforge-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/bevsxyz/WhisperForge/releases/download/v0.5.5/whisperforge-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install bevsxyz/tap/whisperforge

Install prebuilt binaries into your npm project

npm install whisperforge@0.5.5

Download whisperforge 0.5.5

File	Platform	Checksum
whisperforge-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
whisperforge-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
whisperforge-x86_64-pc-windows-msvc.msi	x64 Windows	checksum
whisperforge-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
whisperforge-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

WhisperForge v0.4.0

0.4.0 — 2026-05-22

Breaking — Phase E: foundation refactor (crate merger + CLI UX + device selection).

Removed (breaking)

• whisperforge-convert crate removed — model conversion lives in wf convert.
• whisperforge binary alias dropped — only wf ships.
• --cpu flag on the transcribe path — replaced by --device cpu.
• --task flag on the transcribe path — parsed-then-runtime-errored before; less code, clearer help.
• --wgpu flag (vestigial; was already replaced by feature-driven dispatch).

Added

• wf subcommand surface: transcribe, convert, list-models. No more implicit-default subcommand — must specify one explicitly.
• wf list-models: tabulates .mpk models under the models directory with precision, n_audio_layer, n_mels, file size.
• --device <auto|cpu|wgpu|cuda> on wf transcribe. auto prefers CUDA → WGPU → CPU based on compiled-in features.
• Native CUDA backend via optional burn-cuda 0.21.0 (opt-in --features cuda; requires CUDA toolkit at build time).
• --models-dir <PATH> and WF_MODELS_DIR env var honored by transcribe and list-models.
• Friendly missing-model error pointing at wf list-models and the matching wf convert invocation.

Changed

• whisperforge-cli crate renamed to whisperforge. Workspace shrinks 5 → 4 crates.
• mise release-check smoke-tests --device cuda and the missing-model friendly error.
• Project docs (CLAUDE.md + README + per-crate READMEs) updated for the new CLI surface; Roadmap marks Phase E ✅ COMPLETE and Phase D (WASM) ⏸ DEFERRED. Next: Phase F = streaming realtime.

Migration

Before	After
whisperforge-cli -a foo.wav -m tiny_en	wf transcribe -a foo.wav -m tiny_en
cargo run -p whisperforge-convert -- --output models/tiny_en	wf convert --output models/tiny_en
wf -a foo.wav --cpu	wf transcribe -a foo.wav --device cpu
wf -a foo.wav --wgpu	wf transcribe -a foo.wav --device wgpu (or default auto)
wf -a foo.wav --task translate	(removed — never worked)

Deferred (Phase E carve-outs, queued for later phases)

• VRAM-aware --encoder-batch-size auto-tune. Users override with --encoder-batch-size until a workload justifies the heuristic cost.
• Windows wgpu runtime fallback. Upstream windows-crate conflict between wgpu-hal and gpu-allocator is compile-time; no runtime probe can rescue it. Windows binary still ships CPU-only.

WhisperForge v0.3.1

Changelog

All notable changes to WhisperForge are documented in this file.

Format follows Keep a Changelog.
All workspace crates are versioned together.

---

0.2.0 — 2026-04-30

Added

Core transcription pipeline (whisperforge-core)

• Whisper model architecture for all sizes (tiny.en through large-v2/v3), built on Burn 0.20.
• Audio pipeline: WAV loading with format dispatch (f32/i16/i24/i32), rubato resampling to 16 kHz mono, Slaney-scale mel spectrogram matching Python Whisper (power spectrum, center-padding, 80-band filter bank).
• HybridDecoder with quality-gated temperature fallback matching faster-whisper SOTA: compression ratio gate (2.4), log-probability gate (−1.0), no-speech threshold (0.6), temperature sequence [0.0, 0.2, 0.4, 0.6, 0.8, 1.0].
• KvCache<B> + forward_decoder_cached: O(n) per step via static cross-attention K,V and growing self-attention K,V — ~2.6× speedup over naive O(n²) decoder.
• batch_mel_spectrograms: all audio chunks mel-encoded in a single forward_encoder call before sequential decoding.
• transcribe_with_timestamps: per-token timestamps via cross-attention peak, ~100 ms precision.
• extract_speaker_embedding: mean-pool encoder output + L2-normalise → speaker fingerprint.
• TranscriptionResult / TranscriptionSegment: fully serde-serializable; token_timestamps and speaker fields skipped when absent.
• 0.8% average WER on LJSpeech tiny.en benchmark.

VAD and alignment (whisperforge-align)

• VoiceActivityDetector via earshot with configurable threshold.
• AudioSegmenter: splits audio into voice spans ≤30 s, filters silence.
• BatchedTranscriber<B>: real batch transcription pipeline — batch_mel_spectrograms → forward_encoder → per-segment KV-cached greedy decode → HybridDecoder.
• SrtWriter / SrtEntry: generates well-formed SRT subtitle files.

Speaker diarization (whisperforge-diarize)

• SpeakerDiarizer: agglomerative single-linkage clustering of speaker embeddings.
• cluster_embeddings: cosine similarity with configurable threshold, SPEAKER_NN label assignment.

Model conversion (whisperforge-convert)

• Converts OpenAI Whisper safetensors (via HuggingFace hf-hub) to Burn NamedMpk format.
• Correct tensor name mapping for all encoder and decoder layers including cross-attention.
• Config auto-detection for all model sizes from tensor shapes.

CLI binary (whisperforge / wf)

• Two binary aliases: whisperforge and wf (same binary, shorter alias).
• --audio-file / -a: WAV input.
• --model / -m: model directory under models/ (default: tiny_en_converted).
• --output-format text|srt|json: plain text, SRT subtitles, or JSON with segments and timestamps.
• --decoding-preset fast|balanced|accurate: pre-configured beam size and temperature sequences.
• --beam-size, --temperature, --length-penalty, --no-speech-threshold: per-run overrides.
• --vad-enabled / --vad-threshold: voice activity detection; silence segments skipped.
• --wgpu: WGPU GPU backend (Vulkan, DX12, or Metal — no CUDA required).
• --diarize / --diarize-threshold: speaker labels ([SPEAKER_NN]:) in SRT and JSON.
• --task transcribe|translate, --language: Whisper task and language selection.
• Automatic 30 s chunking with 1 s overlap for long audio.

Notes

• Model files (.mpk, .cfg, tokenizer.json) are not bundled — download from HuggingFace and convert with whisperforge-convert. See README for instructions.
• Requires Rust 1.85+ (Rust 2024 edition).
• --wgpu requires Vulkan/DX12/Metal drivers.
• whisperforge-align has known pre-existing test failures; always excluded from the test suite.
• Phase 7 Option B (ResNet293 speaker embeddings via burn-onnx) is deferred pending Burn 0.21 stable.

v0.3.0

🎯 WhisperForge 0.3.0: Quantization & Performance

🚀 Major Features

Phase C: INT8 Post-Training Quantization ✅

• 4× model size reduction: Tiny (150 MB → 37 MB), ideal for edge deployment
• --quantize int8 flag in model converter
• Transparent loading—no code changes required
• Full precision (FP32) and quantized (INT8) models interoperable

Phase B.5: GPU-Accelerated Mel Spectrogram ✅

• CubeCL DFT kernel for GPU mel filterbank matmul
• --features cubecl-stft enables GPU STFT pipeline
• Faster audio preprocessing on WGPU backend

Burn 0.21 & burn-flex Migration

• Latest Burn with improved numerical stability
• CPU fallback (burn-flex) seamlessly handles CPU inference
• Better WGPU runtime integration

📊 What's Changed

• Quantized models fully compatible with CLI and library API
• Streaming audio pipeline (Phase B) now fully integrated
• Fixed EOT suppression at step 0 for robustness
• Improved error handling across all crates

📦 All 5 Crates Published

• whisperforge-core v0.3.0 — library
• whisperforge-cli v0.3.0 — binary
• whisperforge-convert v0.3.0 — model converter
• whisperforge-align v0.3.0 — VAD + SRT
• whisperforge-diarize v0.3.0 — speaker diarization

🛠️ Dependency Updates

Simplified workspace dependency management: inter-crate deps now use workspace version automatically.

📖 Next Phase

Phase D: WASM Target—browser-native speech-to-text with wasm-bindgen.

Full Changelog: v0.2.0...v0.3.0

WhisperForge v0.2.0

Changelog

All notable changes to WhisperForge are documented in this file.

Format follows Keep a Changelog.
All workspace crates are versioned together.

---

0.2.0 — 2026-04-30

Added

Core transcription pipeline (whisperforge-core)

• Whisper model architecture for all sizes (tiny.en through large-v2/v3), built on Burn 0.20.
• Audio pipeline: WAV loading with format dispatch (f32/i16/i24/i32), rubato resampling to 16 kHz mono, Slaney-scale mel spectrogram matching Python Whisper (power spectrum, center-padding, 80-band filter bank).
• HybridDecoder with quality-gated temperature fallback matching faster-whisper SOTA: compression ratio gate (2.4), log-probability gate (−1.0), no-speech threshold (0.6), temperature sequence [0.0, 0.2, 0.4, 0.6, 0.8, 1.0].
• KvCache<B> + forward_decoder_cached: O(n) per step via static cross-attention K,V and growing self-attention K,V — ~2.6× speedup over naive O(n²) decoder.
• batch_mel_spectrograms: all audio chunks mel-encoded in a single forward_encoder call before sequential decoding.
• transcribe_with_timestamps: per-token timestamps via cross-attention peak, ~100 ms precision.
• extract_speaker_embedding: mean-pool encoder output + L2-normalise → speaker fingerprint.
• TranscriptionResult / TranscriptionSegment: fully serde-serializable; token_timestamps and speaker fields skipped when absent.
• 0.8% average WER on LJSpeech tiny.en benchmark.

VAD and alignment (whisperforge-align)

• VoiceActivityDetector via earshot with configurable threshold.
• AudioSegmenter: splits audio into voice spans ≤30 s, filters silence.
• BatchedTranscriber<B>: real batch transcription pipeline — batch_mel_spectrograms → forward_encoder → per-segment KV-cached greedy decode → HybridDecoder.
• SrtWriter / SrtEntry: generates well-formed SRT subtitle files.

Speaker diarization (whisperforge-diarize)

• SpeakerDiarizer: agglomerative single-linkage clustering of speaker embeddings.
• cluster_embeddings: cosine similarity with configurable threshold, SPEAKER_NN label assignment.

Model conversion (whisperforge-convert)

• Converts OpenAI Whisper safetensors (via HuggingFace hf-hub) to Burn NamedMpk format.
• Correct tensor name mapping for all encoder and decoder layers including cross-attention.
• Config auto-detection for all model sizes from tensor shapes.

CLI binary (whisperforge / wf)

• Two binary aliases: whisperforge and wf (same binary, shorter alias).
• --audio-file / -a: WAV input.
• --model / -m: model directory under models/ (default: tiny_en_converted).
• --output-format text|srt|json: plain text, SRT subtitles, or JSON with segments and timestamps.
• --decoding-preset fast|balanced|accurate: pre-configured beam size and temperature sequences.
• --beam-size, --temperature, --length-penalty, --no-speech-threshold: per-run overrides.
• --vad-enabled / --vad-threshold: voice activity detection; silence segments skipped.
• --wgpu: WGPU GPU backend (Vulkan, DX12, or Metal — no CUDA required).
• --diarize / --diarize-threshold: speaker labels ([SPEAKER_NN]:) in SRT and JSON.
• --task transcribe|translate, --language: Whisper task and language selection.
• Automatic 30 s chunking with 1 s overlap for long audio.

Notes

• Model files (.mpk, .cfg, tokenizer.json) are not bundled — download from HuggingFace and convert with whisperforge-convert. See README for instructions.
• Requires Rust 1.85+ (Rust 2024 edition).
• --wgpu requires Vulkan/DX12/Metal drivers.
• whisperforge-align has known pre-existing test failures; always excluded from the test suite.
• Phase 7 Option B (ResNet293 speaker embeddings via burn-onnx) is deferred pending Burn 0.21 stable.