whisper: word-level timestamps using DTW

THIRD-PARTY-LICENSES.md

@@ -15,6 +15,12 @@ Prebuilt artifacts (the Swift xcframework, the native Python wheels, and the
 npm platform packages) carry these same texts alongside the binaries — see each
 binding's packaging for the bundled `LICENSE.ggml` / `LICENSE.miniz` / `LICENSE.silero_vad` files.
 
+The Whisper alignment-head presets come from
+[whisper.cpp v1.8.3](https://github.com/ggml-org/whisper.cpp/blob/v1.8.3/src/whisper.cpp),
+copyright (c) 2023–2024 The ggml authors, under the
+[MIT license](https://github.com/ggml-org/whisper.cpp/blob/v1.8.3/LICENSE)
+(the terms are reproduced under ggml below).
+
 Conversion-time tooling also includes `scripts/lib/phonon.py`, adapted from
 Fermion Research's Apache-2.0 container reader at
 `FermionResearch/Phonon-2@7e153e4054c0a10db6d47d6fff5c062d4c77c154`.

docs/environment-variables.md

@@ -94,6 +94,7 @@ its var is unset. Convention: `TRANSCRIBE_<FAMILY>_GGUF`.
 | `TRANSCRIBE_COHERE_GGUF` | `cohere_real_smoke`, `cohere_e2e_smoke` |
 | `TRANSCRIBE_GRANITE5_CTC_GGUF` | `granite5_ctc_real_smoke`, `granite5_ctc_e2e_smoke` |
 | `TRANSCRIBE_WHISPER_GGUF` | `whisper_e2e_smoke`, `whisper_tokenize_parity` |
+| `TRANSCRIBE_WHISPER_WORD_GGUF` | `whisper_word_timestamps_smoke` (stock OpenAI Whisper GGUF, e.g. Base) |
 | `TRANSCRIBE_QWEN3_ASR_GGUF` (+ `_0_6B_GGUF` / `_1_7B_GGUF`) | qwen3_asr smokes / parity |
 | `TRANSCRIBE_MOONSHINE_STREAMING_TINY_GGUF` | moonshine_streaming smokes, `stream_offline_interleave_smoke` |
 | `TRANSCRIBE_VOXTRAL_REALTIME_GGUF` | `voxtral_realtime_real_smoke`, `stream_offline_interleave_smoke` |

docs/models/whisper.md

@@ -97,13 +97,26 @@ All Whisper variants support:
 - **Transcription** of 16 kHz mono WAV input.
 - **Long-form audio** via 30-second chunked decoding with the
   prev-context window assembly described in the family doc.
-- **Segment timestamps** — the finest granularity the library emits for
-  Whisper (`max_timestamp_kind = segment`). Word-level timestamps are not
-  currently exposed.
+- **Segment timestamps** — the default under `AUTO`, when no transcript prefix is set.
 - **Translation** (any supported language → English) on multilingual
   checkpoints — `.en` variants are transcribe-only.
 
 What's not supported (consistent across the family): real-time
 streaming (whisper is not streaming-first; chunked 30-second windows
 only), VAD, speaker diarization. See the family doc for the full
 runtime contract.
+
+## Word timestamps
+
+The OpenAI Whisper GGUF variants advertise `TRANSCRIBE_TIMESTAMPS_WORD`
+as their maximum timestamp granularity. Request it with
+`transcribe_run_params::timestamps` or `--timestamps word` in the CLI.
+The decoder aligns the accepted text to the audio with a second cross-attention
+pass and dynamic time warping (DTW). Word times are in milliseconds relative to
+the input, including across 30-second windows. Languages such as Japanese
+use complete UTF-8 token pieces as timing units.
+
+This mode adds decoder work and uses the serial path for offline batches.
+It requires a model-specific alignment-head preset, so legacy `.bin` models
+and other Whisper variants advertise segment granularity. A transcript prefix
+requires `AUTO` or `NONE`.

src/CMakeLists.txt

@@ -97,6 +97,7 @@ add_library(transcribe
     arch/whisper/weights.cpp
     arch/whisper/encoder.cpp
     arch/whisper/decoder.cpp
+    arch/whisper/alignment.cpp
     arch/whisper/bin_load.cpp
     arch/whisper/public.cpp
     arch/moonshine/model.cpp