* metal : init
* whisper : factor out graph builds
* whisper : allocate encoder and decoder using ggml-alloc
* whisper : ggml-alloc is now supported
* whisper : CoreML support ggml-alloc
* build : fix ggml-alloc
* ios : update submodule
* extra : update sync-ggml.sh script to also sync ggml-alloc
* ci : see if this is causing the crash
* whisper : refactor ggml-alloc init
* whisper.android : try to fix build
* whisper : initial Metal version
* ci : try to debug vmem issue
* metal : decoder works on GPU!
* metal : add multi-decoder support
* ggml : fix ggml_nbytes (probably temp solution)
* metal : run "cross" step on the GPU
* whisper : remove ggml_repeat in the encoder
* whisper : offload the Encoder to Metal
* ggml : use simpler ggml_bytes() implementation
* ggml-alloc : try to make CI happy by reducing vram to 128GB
* whisper : add whisper_allocr to wrap ggml_allocr
* whisper : factor out alloc init in a function
* cmake : update to support Metal build
* whisper : add <functional> header
* objc : fix build (no Metal yet)
* ios : add Metal support
* swiftui : fix build
* metal : speed-up KQ multiplication
* metal : sync latest llama.cpp kernels
* readme : add Metal info
* ios : update submodule
* coreml : add code to toggle Core ML config (CPU, ANE, GPU)
* bench : fix timings by running a pre-heat
* bench : start benching the decoder
* whisper : add ggml_mul_mat_pad
* bench : fix uninitialized vars
* whisper : add comment for disabling mul-mat padding
* whisper : add description of ggml_mul_mat_pad
* whisper : clean-up ggml_mul_mat_pad
* metal : remove the "concurrent" flag
* bench : variable n_past
* ios : update SPM package
* Do not use _GNU_SOURCE gratuitously.
What is needed to build whisper.cpp and examples is availability of
stuff defined in The Open Group Base Specifications Issue 6
(https://pubs.opengroup.org/onlinepubs/009695399/) known also as
Single Unix Specification v3 (SUSv3) or POSIX.1-2001 + XSI extensions,
plus some stuff from BSD that is not specified in POSIX.1.
Well, that was true until NUMA support was added recently in ggml,
so enable GNU libc extensions for Linux builds to cover that.
There is no need to penalize musl libc which simply follows standards.
Not having feature test macros in source code gives greater flexibility
to those wanting to reuse it in 3rd party app, as they can build it with
minimal FTM (_XOPEN_SOURCE=600) or other FTM depending on their needs.
It builds without issues in Alpine (musl libc), Ubuntu (glibc), MSYS2.
* examples : include SDL headers before other headers
Avoid macOS build error when _DARWIN_C_SOURCE is not defined, brought by
SDL2 relying on Darwin extension memset_pattern4/8/16 (from string.h).
* make : enable BSD extensions for DragonFlyBSD to expose RLIMIT_MEMLOCK
* make : use BSD-specific FTMs to enable alloca on BSDs
* make : fix OpenBSD build by exposing newer POSIX definitions
* cmake : follow recent FTM improvements from Makefile
* Initial proof of concept Vim plugin
At present, this is likely only slightly better than feature parity with
the existing whisper.nvim
Known issues:
Trailing whitespace
Up to an existing length(5 seconds) of speech may be processed when
listening is enabled
CPU cycles are spent processing speech even when not listening.
Fixing these issues is likely dependent upon future efforts to create a
dedicated library instead of wrapping examples/stream
* Support $WHISPER_CPP_HOME environment variable
A minor misunderstanding of the whisper.nvim implementation resulted in
a plugin that was functional, but not a drop in replacement as it should
be now.
* Initial progress on LSP implementation
Libcall is nonviable because the library is immediately freed after a
call is made. Further investigation has shown Language Server Protocol as
a promising alternative that both simplifies the required logic on the
vimscript side and increases the ease with which plugins for other
editors could be made in the future. This is a very large undertaking
and my progress has slowed substantially.
Work is far from being in a usable state, but I wish to keep track of
major refactors for organizational purposes.
* Rewrite audio windowing of guided transcription
One of the defining goals of this venture is allowing consecutive
commands to be rattled off without the existing deadzones of the current
implementation.
* Add unguided_transcription. Cleanup.
The unguided transcription implantation heavily borrows from existing
example implementations and the guided_transcription logic.
A high level pass was done to check that method arguments are accurate
to what inputs are actually required.
A first attempt at cancellation support was added for record keeping,
but will be deleted in a future commit.
* Fix compilation.
Resolves a large number of compilation errors.
No testing has been done yet for execution errors.
Update Makefile and .gitignore
* Functional unguided_transcription
* Functional guided_transcription
Fix commandset_list being passed by value
Properly register the first token of a multitoken command
* Minor changes before time fix
I've apparently made an awfully major mistake in thinking that unix time
was in milliseconds and will be changing all timekeeping code to use
standardized methods.
In preparation for this is a number of minor bugfixes.
Output is manually flushed.
An echo method has been added.
registerCommandset now wraps the returned index
* Swap timekeeping to use std::chrono
* Add work in progress lsp backed whisper.vim plugin
Current progress blockers are
Adding modality awareness to the command processing
(specifically, motion prompting)
Improving the VAD to be a little more responsive
(testing start of activity)
* Reworked vim plugin command loop
* Fix change inside
Multiple bug fixes that, crucially, bring the plugin to the point where a
demonstration video is possible
Add better echo messaging so whisper_log isn't required
Add loading complete message as indicator when listening has started
Insert/append are actually included in command sets
Some more heavy handed corrections to prevent a double exit when leaving
insert mode
As a somewhat hacky fix, the very first space is removed when inserting.
This cleans up most use cases, but leaves me unsatisfied with the few
cases it would be desired.
* Forcibly set commandset_index to 0 after subinsert
Also remove unnecessary ! to use builtin vim command
* Fix upper
A minor scope mistake was causing upper'd inputs to be eaten.
This was fixed and echoing was slightly improved for clarity.
* Fix formatting
Corrects indentation to 4 spaces as project standard
Slightly better error support for malformed json input
* Remove obsolete vim plugin
* Add json.hpp library
The same library that is used for the llama.cpp server
* Minor cleanups
add lsp to the make clean directive.
remove a redundant params definition.
reorder whisper.vim logging for subtranscriptions
Corrections to unlets (variables of argument scope appear immutable)
* Fix indentation. Fallback for subTranscription
Indentation has been changed to 4 spaces.
Unit testing has been set up, I'm opting not to include it in the
repository for now.
It however, has revealed a bug in the state logic where a
subtranscription can be initiated without having a saved command
When this occurs, append is added as a fallback
* Move audio polling logic to a subfunction
While work on the improved vad will continue, It's grown to be a little
out of scope. Instead, a future commit will perform multiple detection
passes at substretches of audio when a backlog of audio exists.
To facilitate this, and prevent code duplication, the vad code has been
moved into a subfunction shared by both the unguided and guided
transcription functions.
* Test for voice over subchunks if backlog > 1s
As the existing VAD implementation only checks for a falling edge at the
end of an audio chunk. It fails to detect voice in cases where the
recorded voice is only at the beginning of the audio.
To ameliorate this, when the timestamp would cause analysis of audio
over a second in length, it is split into 1 second length subchunks
which are individually tested.
Results are promising, but there seems to be a remaining bug with
unguided transcription likely related to saving context
* Limit the maximum length of audio input.
This existing VAD implementation only detects falling edges, which
means any gap in the users speaking is processed for transcription.
This simply establishes a constant maximum length depending on the type
of transcription. Uguided gets a generous 10 seconds and guided, 2.
While quick testing showed that commands are generally around a half a
second to a second, limiting commands to an even second resulted in
extreme degradation of quality. (Seemingly always the same output for a
given commandset)
* Unguided timestamp tracking, cleanup
Unguided transcriptions where not setup to allow for passing of
timestamp data forward, but have been corrected.
No_context is now always set to false. While conceptually desirable for
the quality of guided transcription, It was seemingly responsible for
prior command inputs ghosting in unguided transcription.
Save and Run are now tracked by command number instead of command text.
While command_text was provided for convenience, I wish to keep command
index authoritative. This gives greater consistency and potentially
allows for end users to rename or even translate the spoken versions of
these commands
* By default, maintain mode.
Previously, mode was reset to 0 unless otherwise set.
In addition to causing some edge cases, this was didn't mesh well with
the existing approach to visual mode.
With this change, initial tests indicate visual mode is functional.
* Add undo breaks before subtranscriptions
Subtranscriptions use undo as a hack to allow for partial responses to
be displayed. However, scripts don't cause an undo break mid execution
unless specifically instructed to. This meant that multiple
unguided transcriptions from a single session would cause a latter to
undo a former.
This is now fixed and undo should be reasonably usable as a command.
* Append instead of insert for new undo sequence
When entering and leavening insert mode with `i`, the cursor shifts one
column to the left. This is remedied by using append instead of insert
for setting these breaks in the undo sequence
`-` was also added to the pronunciation dictionary to be pronounced as
minus as it was causing a particularly high failure rate.
* Move undo sequence breaks to command execution
Previously, undo sequence breaks were triggered when there was a command
that caused a move to insert mode. This caused commands that changed
state (like delete or paste) to be bundled together with into the last
command that caused text to be entered.
* Fix repeat. Add space, carrot, dollar commands
Repeat (.) wasn't being tracked properly just like undo and is being
manually tracked now.
While efforts have been made to properly handle spaces, it was
particularly finicky to add a single space when one is needed. A
special 'space' command has been added to insert a single space and move
the cursor after it.
Carrot and Dollar commands have been added for start of line and end of
line respectively. These are both simple to implement, and just a
matter of defining a pronunciation.
* Return error on duplicate in commandset
Not every command in the commandset tokenizes to a single token.
Because of this, it's possible for that two commands could resolve to
the same single token after subsequent tokens are discarded.
This commit adds a simple check for duplicates when a commandset is
registered and returns an error if so.
Additional code will be required later on the vim side to actually
process this error.
* Add support for user-defined commands
This adds a user definable dictionary from spoken keys to strings or
funcrefs. All keys are added to the commandlist and when spoken, trigger
the corresponding function.
Like "save" and "run", these user commands are only available when the
command buffer is empty.
* Add readme, update cmake
* Add area commandset. Refactor spoken_dict
Area commands (inside word, around sentence...) have been given a
commandset as considered earlier.
Verbose definitions for spoken_dict entries now use dicts instead of
lists. This shortens the definition for most keys that require it and
scales better with the addition of further commandsets
* Add mark, jump. Fix change under visual.
Mark (m) and jump (') have been added.
When a visual selection was executed upon a command that initiated a
subtranscription (change) the area of the visual selection is not
properly tracked which causes the attempt to stream in partial response
to fail. This is solved by disabling partial transcriptions from being
streamed when a subtranscription is started while in visual mode.
* Accommodate ignorecase. Fix change.
From testing on older different versions of vim, the test for
distinguishing an 'R' replace all from an 'r' replace could fail if
ignorecase was set. The comparison has been changed to explicitly
require case matching
Change detection has been moved to the execution section as it was missing the
change+motion case.
* Support registers. Fix README typo
There's no logic to prevent doubled register entry, but the functional
result is equivalent to if the same key order was typed into vim.
A minor typo in the readme. I've mismemorized the mnemonic for 't' as 'to'
instead of till., but 'to' can't be used as it's a homophone with '2'.
While there was no mistake in the actual logic, it was misleading to use
'to' in the readme.
* 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>
Current `progress_step` was hardcoded into whisper.cpp, this resulted in
bindings having to access progress only at that step even if progress
callback was being called at every iteration.
With this change we get greater granularity progress reporting from
whisper.cpp and bindings/implementations can define their own progress step.
* Initial proof of concept Vim plugin
At present, this is likely only slightly better than feature parity with
the existing whisper.nvim
Known issues:
Trailing whitespace
Up to an existing length(5 seconds) of speech may be processed when
listening is enabled
CPU cycles are spent processing speech even when not listening.
Fixing these issues is likely dependent upon future efforts to create a
dedicated library instead of wrapping examples/stream
* Support $WHISPER_CPP_HOME environment variable
A minor misunderstanding of the whisper.nvim implementation resulted in
a plugin that was functional, but not a drop in replacement as it should
be now.
* 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>
* talk-llama : use posix_madvise() instead of madvise() derived from BSD
sed -i 's,\<madvise\>,posix_&,g;s,\<MADV_,POSIX_&,g' examples/talk-llama/llama-util.h
* make : enable Darwin extensions for macOS builds
This is an attempt at fixing macOS build error coming from the fact that
RLIMIT_MEMLOCK define is not available there without Darwin extensions.
* Do not use _GNU_SOURCE gratuitously.
What is needed to build whisper.cpp and examples is availability of
stuff defined in The Open Group Base Specifications Issue 6
(https://pubs.opengroup.org/onlinepubs/009695399/) known also as
Single Unix Specification v3 (SUSv3) or POSIX.1-2001 + XSI extensions.
There is no need to penalize musl libc which simply follows standards.
Not having feature test macros in source code gives greater flexibility
to those wanting to reuse it in 3rd party app, as they can build it with
minimal FTM (_XOPEN_SOURCE=600) or other FTM depending on their needs.
It builds without issues in Alpine (musl libc), Ubuntu (glibc), MSYS2.
* examples : include SDL headers before other headers
This is an attempt at fixing macOS build error coming from SDL2 relying
on Darwin extension memset_pattern4/8/16 coming from Apple's string.h.
* 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
* feat: adding session support
* readme: adding --session info in examples/talk-llama
* llama: adding session fixes
* readme: updating session doc
* talk-llama: update the value of need_to_save_session to true in order to save the session in the subsequent interaction
* talk-llama: adding missing function which updates session_tokens
Updated the escape_double_quotes() function such that the function now escapes both double quotes and backslashes in the input string.
Changes Made:
- Renamed the function to escape_quotes_and_backslashes
- Modified the condition in the first loop to increment the value of 'escaped_length' for both double quotes and backslashes.
- Modified the condition in second loop to add a backslash before the current character if it is a double quote or a backslash.
Resolves: #769
I disabled this because there were many complaints about slow decoding.
The current implementation does not allow batching the decoders when
using the "best of" or "beam size" parameters, so the decoding time is
proportional to the number of decoders, which is obviously not great.
However, now there are even more complaints about wrong decodings and
repetition.
So, making a compromise by re-enabling the fallbacks, but defaulting to
just 2 "best of" / "beam size" decoders. Also, the temperature step is
increased from 0.2 to 0.4 - i.e. from maximum of 5 fallbacks to maximum
of 2.
Also, the stream example now has fallbacks enabled by default.
close#471#477#508#612#719#731
There is `speak.sh` file in `./examples/talk-llama` as described in README.
However `./examples/talk/speak.sh` is used in `talk-llama.cpp`, this commit corrects that.
* fixed blocking code on node addon
* modify the example to run async
* format
* added logic to see the whisper output
* added logic to see the whisper output
* removed extra function for more clean example
* fixed whisper test to new async implementation
* fixed blocking code on node addon
* modify the example to run async
* format
* added logic to see the whisper output
* added logic to see the whisper output
* removed extra function for more clean example
* 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>
* examples : refactor common code into a library
* examples : refactor common SDL code into a library
* make : update Makefile to use common libs
* common : fix MSVC M_PI ..
* addon.node : link common lib
* addon: implement node addon call whisper through cpp
* addon: modify the license to MIT
* addon: remove iostream
* addon: rename dir
* addon: fix typo
* addon: configure cmake to build when cmake-js is used
Adds -ocsv, aka --output-csv feature to examples/main, which outputs a CSV file containing lines formatted as follows <startTime-in-integer-milliseconds>, <endTime-in-integer-milliseconds>, "<transcript-line-including-commas>".
- remove unnecessary initialization of string to ""
- use empty() instead of checking size()
- use emplace_back instead of push_back
- use nullptr instead of NULL
- remove unnecessary call to .data() on string
- use character overload of find_first_of() instead of passing a string
Android armeabi-v7a's NEON support doesn't support FMA unless configured with `-mfpu=neon-fp-armv8`, which would need runtime checks.
* Also removed ABI filter from Android project.
* 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>
* Add Android sample
* Use main project C files
* Stop existing playback before starting new playback
* Make text scrollable
* Stop playback when starting to record
* Remove extra var
* npm : preparing infra for node package
* npm : package infra ready
* npm : initial version ready
* npm : change name to whisper.cpp
whisper.js is taken
The old "large" model is now renamed "large-v1".
If you have been using it, make sure to rename it and download the new
"large" model for best results.
Same as the command-line tool "command", but runs in the browser
Also, added helper script "extra/deploy-wasm.sh" and fixed some timing
constants for the WASM examples.
* 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.
