* whisper : migrate to ggml-backend
* whisper : fix logit reading
* whisper : fix tensor allocation during load
* whisper : fix beam-search with CUDA
* whisper : free backends + fix compile warning
* whisper : print when CUDA is enabled
* whisper : fix CoreML
* make : clean-up
* talk : fix compile warning
* whisper : support ggml_conv with CUDA and Metal (#1473)
* ggml : add CUDA support for ggml_conv
* whisper : remove ggml_repeat for conv bias + single backend
* cuda : fix im2col kernel
* metal : add im2col support + mul mat-vec f16 x f16
* bench-all : add q4 models
* whisper : clean-up
* quantize-all : fix
* ggml : im2col opts
* whisper : avoid whisper_model_data wrapper
* whisper : add note that ggml_mul_mat_pad does not work with CUDA
* whisper : factor out graph compute in common function
* whisper : fixes
* whisper : fix UB with measure buffers
* whisper : try to fix the parallel whisper_state functionality (#1479)
* whisper : try to fix the parallel whisper_state functionality
* whisper : fix multi-state Metal
* whisper : free backend instances in whisper_state
* whisper : check state->ctx_metal not null
* whisper : add whisper_context_params { use_gpu }
* whisper : new API with params & deprecate old API
* examples : use no-gpu param && whisper_init_from_file_with_params
* whisper.objc : enable metal & disable on simulator
* whisper.swiftui, metal : enable metal & support load default.metallib
* whisper.android : use new API
* bindings : use new API
* addon.node : fix build & test
* bindings : updata java binding
* bindings : add missing whisper_context_default_params_by_ref WHISPER_API for java
* metal : use SWIFTPM_MODULE_BUNDLE for GGML_SWIFT and reuse library load
* metal : move bundle var into block
* metal : use SWIFT_PACKAGE instead of GGML_SWIFT
* style : minor updates
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix MSVC compile error C3688
Instead of simply using 'add_compile_options(/utf-8)' to address the MSVC compile error C3688, a better approach would be to handle it in a way that prevents passing '/utf-8' to NVCC.
* Significantly improve inference quality
In the function `log_mel_spectrogram_worker_thread`, there's an array out-of-bounds issue occurring during the calculation of complex number moduli. This issue is causing disruptions in the FFT spectrum, which, in turn, is reducing the quality of inference.
* Significantly improve inference quality
At last, I've pinpointed the actual source of the problem. Given that the frequency spectrum generated from real input data is symmetrical around the Nyquist frequency, there's a for-loop within the `log_mel_spectrogram_worker_thread` function that attempts to fold the frequency spectrum. Regrettably, a bug within this for-loop is causing a frame shift in the frequency spectrum. The previous attempt to remedy this, which involved using `fft_size + 1` when calculating the modulus, was merely a band-aid solution and did not address the underlying issue.
* Addressed a few minor issues
Fixed the issue of `fft_out` continuously expanding. Resolved the fallback caused by using 'break' instead of `fft_in[j] = 0`.
* Significantly improve inference quality
Thanks for your patience everyone. It's finally sorted out. Now, the right side of the FFT spectrum is being flipped over to the left, and the amplitudes at corresponding positions on the left and right are added together (the spectrum on the left needs to be shifted by one position), then the average is calculated. FFT_OUT[0] is no longer discarded, making full use of the limited space to pack in more information.
* Add annotation and performance improvement
* Calculate FFT only when fft_in are not all zero
* Some minor performance improvement
* Fixed a bug impacting inference quality
* The first version after all the analysis is completed.
* Fix some bugs and add debug mode
* Fixed several bugs
* Temporarily disable speed-up mode and add debug mode.
* Add debug mode
* Disable speed-up mode and add debug mode
* Fix CI error (#1)
* Fix error
* Fix error
* Fixed several bugs including [BLANK_AUDIO] problem
* Remove Hard-coded hann window
* Some Final Fix (#2)
* Fix error
* Fix error
* Probably the last commit
* Probably the last commit
* whisper : minor coding style changes
* whisper : remove debug from public API
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* expose api to let user control log output
Add
whisper_set_log_callback()
that lets user set a callback for log messages.
Change all the
fprintf(stderr, ...)
to call via the above.
* whisper : add <cstdarg>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* add HuggingFace mirror to download ggml model
* support tdrz via simple hack overriding solm tokens
* fix incorrect translate/transcribe token_ids that are not static const
* add apollo 13 sample for tdrz demo
* render [SPEAKER TURN] consistently in all terminal output using vocab.id_to_token
* extend whisper_segment with speaker_turn_next field and save in json output
* fix failing go build
* slipped in some python syntax whoops
* whisper : finalize tinydiarize support (add flag + fixes)
* whisper : tdrz support for word-level timestamps (respect max_len)
* java : try to fix tests after adding tdrz_enable flag
* main : remove TODO leftover
* java : fix params order list after adding "tdrz_enable"
* whisper : fix solm and add nosp token
* main : print tinydiarize help
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Java needs to call `whisper_full_default_params_by_ref()`, returning struct by val does not seem to work.
* added convenience methods to WhisperFullParams
* Remove unused WhisperJavaParams
* add detectlanguage flag
* renaming and help
* no idea why that last one didn't commit
* run language detection if dl is set
* help message fix
* various fixes
* fix quitting
* fix language being english on print
* examples : provide option for exporting also as JSON file (ggerganov/whisper.cpp#614)
* main : remove leftovers
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* whisper.android: Support benchmark for Android example.
* whisper.android: update screenshot in README.
* update: Make text selectable for copy & paste.
* Update whisper.h to restore API name
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* whisper.android: Restore original API names.
---------
Co-authored-by: tinoue <tinoue@xevo.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Added whisper state + default state on the whisper_context
* Fixed some examples and bindings
* Fixed whisper_n_len (which was used in some binding) and added whisper_n_len_from_state
* Fixed comments
* whisper : reuse kv_cache_free() and fix compiler warnings
* whisper : clean-up the API comments
---------
Co-authored-by: Sandro Hanea <sandrohanea@microsoft.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Resolves some of warnings when compiling with clang/clang++
Mostly nit stuff that clang catches when compiling with -Wall -Wextra
-pedantic.
- Fix comparison between sign/unsigned integers.
- Passes a constant reference (const&) instead of copying each time.
* minor : normalize coding style
* minor : fix warning
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* whisper : try to improve the token sampling strategy
- Add the "max_initial_timestaamp" token logic from OpenAI
- Disallow sampling timestamps that are in the past
* whisper : fix the max initial timestamp logic + fallback decoding
* feat: prompt previous tokens for streaming
I used a vector pointer instead of vector itself because it gave weird errors, and why not
* convert vector to use with C api
* feat: remove old refs, check for prompt size
* feat: use better way of getting the pointer
Used to overwrite the audio context size of the Encoder.
For example, setting "audio_ctx = 512" will make it run about 3 times
faster, processing about 10s of audio, instead of 30s.
The transcription quality drops, but this can be used for real-time
streaming purposes where performance is important.
Using a Phase Vocoder for speeding up the audio tempo by scaling down
the frequencies in the frequency domain.
This reduces the computation in the Encoder by a factor of 2.
The transcription accuracy is degraded, but for slow to normal speech -
it seems to be still very good.
I think this can find application for real-time transcription - i.e. the
"stream" example.
Georgi Gerganov
Jhen-Jie Hong
mkiol
Didzis Gosko
bobqianic
Fangjun Kuang
Yang Shen
Evan Martin
Ryan Metcalfe
Akash Mahajan
Nicholas Albion
CRD716
GitAritron
pajowu
Leo Moll
Takeshi Inoue
sandrohanea
shibukazu
kamranjon
Matija Pevec
Syahmi Azhar
Matheus de Sousa
M. Eren Akbiyik