Fix --vad-min-silence-duration-ms (-vsd) parsing in whisper-vad-speech-segments.
The option was incorrectly assigned to vad_min_speech_duration_ms instead of vad_min_silence_duration_ms. As a result, the requested silence duration was ignored and the minimum speech duration was overwritten.
* Fix#587: Remove leading space from txt output
The BPE tokenizer used by Whisper produces tokens with leading spaces,
causing each line in the txt output to start with an unwanted space.
This fix strips leading whitespace (spaces and tabs) from each segment
when writing to txt output files, improving the readability of the
transcription output.
Fixes: https://github.com/ggml-org/whisper.cpp/issues/587
Change the empty --hf-file branch of whisper_hf_resolve_model to be
cache-first and refuse ambiguity rather than pick the repo's first
ggml-*.bin. With no -hff: exactly one cached ggml-*.bin resolves it
(no network); more than one errors and lists the cached files; a cold
cache errors and lists the repo's available models instead of silently
downloading. This makes "download once with -hff, then just -hf" work.
Unlike llama.cpp's -hf <user>/<model>[:quant] default-quant pick
(find_best_model), whisper repos are many-models-one-repo with no
meaningful default, so we key off the cache and error+list on ambiguity.
The explicit -hff path (download-first, cache fall-back) is unchanged.
whisper_hf_resolve_model now prints a specific diagnostic for every
failure mode, so cli.cpp no longer prints its own generic (and now
inaccurate) "not found in HF cache" line; it just returns exit 3.
tests/test-hf-resolve.sh gains single-cached (-hf alone -> exit 0) and
multi-cached (-hf alone -> exit 3 + "multiple models cached" + list)
cases, and the missing-file assertion matches the new message.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Enable the network half of the HF cache resolver. whisper_hf_resolve_model
now (unless HF_HUB_OFFLINE is set) lists the repo over the HF API, picks the
primary file (exact --hf-file, else first ggml-*.bin), downloads it into the
blobs/<oid> + snapshots/<commit> layout via a slim hf_cache::download_file,
and falls back to the Phase 1 on-disk cache scan on empty listing or network
failure. HF_TOKEN is honored for the Authorization: Bearer header.
HTTPS is gated behind a new WHISPER_OPENSSL CMake option (find_package OpenSSL,
CPPHTTPLIB_OPENSSL_SUPPORT, link OpenSSL::SSL/Crypto); an https attempt in a
non-SSL build prints the rebuild hint.
download_file follows redirects manually and disables httplib url-encoding:
cpp-httplib 0.20 (whisper's vendored version) both mishandles cross-host
redirects and re-encodes the already-encoded signed xet CDN URL, corrupting
the presigned query string into a 403. The bearer token is dropped on
cross-host (CDN) redirects.
tests/test-hf-resolve.sh gains an offline fall-back case (HF_HUB_OFFLINE=1),
a WHISPER_CLI override, and an optional no-OpenSSL rebuild-hint check.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Port llama.cpp's HuggingFace hub-cache subsystem (http.h, hf-cache.{h,cpp})
into whisper.cpp's shared common library and wire -hf/--hf-repo +
-hff/--hf-file into whisper-cli. Phase 1 is cache-only: whisper_hf_resolve_model
scans the on-disk HF hub cache (get_cached_files + finalize_file) and maps
org/repo (+ optional file) to a concrete snapshot path, so a model already
pulled by the hf CLI resolves with no -m path. An explicit -m still wins and
the no-args default stays models/ggml-base.en.bin.
The network download path is compiled but unused this phase (enabled in
Phase 2). Adds tests/test-hf-resolve.sh covering cache hit, missing-file
error (exit 3), -m regression, and the default path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* ci, examples : add package-lock.json to examples addon node
This commit commits the package-lock.json file for examples/addon.node.
This is needed to make the CI workflow work properly and after we
started pinning the action versions in the workflow, the CI started
failing because the package-lock.json was not committed. At least that
is what I think is happening.
* ci: use node 20.x and 22.x instead of 16.x and 18.x
* fix CMakeLists.txt target properties for addon.node
* fix/android-debug-inference-issue: Applied BUILD_TYPE=Release to NDK build of whisper.android example
* fix/android-debug-inference-issue: Applied BUILD_TYPE=Release to NDK build of whisper.android.java example
Fixed the -vspd flag for vad_min_speech_duration_ms, to prevent hiding vad_min_silence_duration_ms.
In usage () clarified the output timestamp units.
Fixed a few typos.
Adds a `--version` option to whisper-cli that prints the library version
via `whisper_version()` and exits, plus a corresponding entry in the help
output. Mirrors the existing `-h`/`--help` handling.
Closes#608
* cli : merge tokens split across UTF-8 boundaries in JSON output
When a multi-byte UTF-8 codepoint (most commonly a CJK character, 3 bytes)
is split across multiple whisper tokens, the -ojf/--output-json-full
writer emitted each token's partial bytes as its own JSON string, producing
invalid UTF-8 that chokes downstream parsers.
Merge adjacent tokens in output_json whenever the accumulated text still
ends on an incomplete UTF-8 sequence. The merged entry keeps the first
token's id/p/t_dtw and extends t1 to the last absorbed token, which
matches how segment text is assembled elsewhere.
Refs #1798
* fix: address review — add braces for consistency, use full issue URL
- Add braces to if/else chain for codebase consistency
- Use full URL for issue #1798 reference
Review: @danbev
---------
Co-authored-by: texasich <texasich@users.noreply.github.com>
Co-authored-by: texasich <texasich@gmail.com>
* common: add memory buffer overload of read_audio_data
whisper-server /inference without --convert passed the uploaded file
bytes to read_audio_data as a filename, so ma_decoder_init_file tried
to open a path starting with "RIFF" and failed. every request returned
HTTP 400 "Invalid request" on builds without WHISPER_FFMPEG, which is
the default.
factor the PCM extraction into a shared helper and add an overload that
decodes straight from a memory buffer via ma_decoder_init_memory, which
the function already used for the stdin path. server now calls it with
the upload content. the filename overload behavior is unchanged.
This commit addresses a memory leak in the `read_audio_data` function
where it is currently possible that a call to `ma_decoder_init_file`
succeeds and the function returns early without calling
`ma_decoder_uninit`. A similar situation can occur with
`ma_decoder_init_memory`.
Refs: https://bugs.debian.org/1124796
Co-authored-by: Daniel Bevenius <daniel.bevenius@gmail.com>
The project moved from ggerganov/ to ggml-org/ and the README already
references the new URL in both places it mentions issue #89 (README.md
and examples/bench/README.md). Syncing the two remaining hardcoded URLs
in examples/bench/bench.cpp and examples/bench.wasm/emscripten.cpp.
The old URL still redirects, so this is cosmetic.
Several error paths in the /inference and /load endpoints returned
HTTP 200 with a JSON error body, making it impossible for clients
to distinguish errors from successful responses by status code.
Set 400 for client errors (missing file field, unreadable audio,
missing/invalid model) and 500 for server errors (ffmpeg conversion
failure). The two existing status-code sites (499 for client
disconnect, 500 for processing failure) are unchanged.
djd0723
Georgi Gerganov
Igor Koshenskii
Cappuccino
Dominic Dones
Daniel Bevenius
Yongmin Yoo 유용민
QuantiusBenignus
Rum Nguyen
Noah Lyons
texasich
Pascal
petterreinholdtsen
Andreas Lubbe
jinweihan
Gaël James
Igor Loskutov
Dmitry Atamanov
Sid Mohan