NEWS

---

whisper 0.4.0 (2026-06-20)

• New serve(): a single-process, OpenAI-compatible HTTP STT server (POST /v1/audio/transcriptions and /translations, GET /health) built on base R sockets, with no new dependencies. It loads the model once and keeps it resident, so it drops in for the OpenAI API or a Whisper container; point stt.api at it with set_stt_base(). Returns text, json, or verbose_json (segment timestamps, plus per-word timestamps when the request includes timestamp_granularities[]=word). An example systemd unit ships in system.file("whisper.service", package = "whisper").
• JIT decoding on CUDA: each generated token's decoder forward runs as one jit_compile'd TorchScript call instead of dozens of dispatched R->torch calls, several times faster end-to-end and token-for-token equivalent to the eager path. Covers both greedy and word-timestamp decoding. On by default via the new jit argument to transcribe()/whisper_pipeline(); pass jit = FALSE for the eager decoder. No effect on CPU or beam search.
• Silence handling now matches the reference Whisper, fixing transcripts that ran past the end of short audio. Three changes, ported from openai-whisper: decoding suppresses non-speech tokens (brackets, music notes, speaker tags) and control tokens at every step, so output no longer contains [BLANK_AUDIO]/[MUSIC PLAYING]-style annotations; the seek loop decodes only the real audio (content_frames), not the fixed 30s of mel padding, so a 7s clip no longer trails off into hallucinated text up to 30s; and a no-speech-probability gate skips windows that read as silence. The special-token table gains sot_lm and sot_prev.
• Bound and mitigate degenerate repetition loops, matching the reference. A long non-speech sound (e.g. a laugh) could make the decoder emit one token ("ha") hundreds of times - garbage output, and enough accumulated cross-attention to exhaust memory on a small GPU. Decoding is now capped at half the text context (the reference's sample_len) rather than the full context, and the default temperatures enable the existing compression-ratio fallback, which re-decodes too-repetitive output at a higher temperature.
• Fix tokenizer_encode() crashing for models whose vocab.json omits the <|endoftext|> key (large-v3): the end-of-text id now comes from the special-token table (as in the Python reference, which keeps special tokens out of the BPE vocab), and the lookup can no longer return a list. A regression test covers a vocab without the key, and encode_special() resolves the core special tokens from the table too.
• whisper_dtype() now falls back to float32 on the GTX 16-series (TU116/TU117: GTX 1630/1650/1660 and Ti/Super variants), which compute fp16 incorrectly and return NaN (seen as repeated "!" tokens). Detection is by GPU name, CUDA-gated and tryCatch-guarded (dormant on non-CUDA/CRAN machines); pass dtype = "float16" to override.
• New whisper_tune_gc(): opt-in helper that tunes torch's CUDA allocator GC rates for inference. No-op off CUDA, and only sets options that are unset.
• Scaled dot-product attention now calls the exported torch::torch_scaled_dot_product_attention() instead of reaching into torch's namespace; the torch dependency is floored at 0.17.0, where it is exported.
• README performance table refreshed for the JIT word-timestamp path.

whisper 0.3.0 (2026-03-13)

• Language auto-detection: transcribe() now defaults to language = NULL, which detects the spoken language from the audio before decoding. New exported function detect_language() for standalone language identification. Breaking: previous default was language = "en". Code relying on the default now auto-detects instead of assuming English. Pass language = "en" explicitly to restore old behavior.
• Segment-level and word-level timestamps via DTW alignment
• Beam search decoding with temperature sampling and fallback
• SDPA attention (FlashAttention on GPU)
• whisper_pipeline() for cached model reuse across multiple transcriptions
• Hardcoded special token table (eliminates added_tokens.json download)
• Fixed invalid multibyte string crash in BPE decoder
• Fixed DTW boundary guards and seek loop in transcribe_chunk()

whisper 0.1.0 (2026-02-06)

• Initial CRAN submission
• Native R torch implementation of OpenAI Whisper
• Support for all model sizes: tiny, base, small, medium, large-v3
• Automatic model download from HuggingFace
• Model-specific special token handling for large-v3 compatibility
• KV caching for efficient autoregressive decoding
• Long audio chunking for files longer than 30 seconds
• Optional timestamp and segment extraction