Vosklet/Documentation.md

API Reference

JS window object

Function/Object	Description
Promise<Module> loadVosklet()	Load Vosklet module interface

Module object

Function/Object	Description
Promise<Model> createModel(url: string, path: string, id: string) Promise<SpkModel> createSpkModel(url: string, path: string, id: string)	Create a Model or SpkModel, model files must be directly under the model root, and compressed model must be in .tar.gz/.tgz format. Tar format must be USTAR. If: - path contains valid model files and id is the same, there will not be a fetch from url. - path doesn't contain valid model files, or if it contains valid model files but id is different, there will be a fetch from url, and the model is stored with id. Models are thread-safe and reusable across recognizers. Compressed .tar.gz models are automatically decompressed before caching. The browser cache stores validated uncompressed USTAR tar archives internally. Invalid cache entries are removed automatically.
Promise<Recognizer> createRecognizer(model: Model, sampleRate: float) Promise<Recognizer> createRecognizerWithSpkModel(model: Model, sampleRate: float, spkModel: spkModel) Promise<Recognizer> createRecognizerWithGrm(model: Model, grammar: string, sampleRate: float)	Create a Recognizer
setLogLevel(lvl: int)	Set log level for Kaldi messages (default: 0: Info) -2: Error -1: Warning 1: Verbose 2: More verbose 3: Debug
Promise<AudioWorkletNode> createTransferer(ctx: AudioContext, bufferSize: int)	Create a node that transfer its inputs back to the main thread with custom buffer size (must be multiple of 128). Its port's onmessage handler can be set to get audio data. Has 1 input with 1 channel and no output. The the higher the size, the lesser the audio breaks up, but the higher the latency. Recomended value is around 128 * 150.
Promise<void> cleanUp()	A convenience function that call delete() on all objects and revoke all URLs. Run this when you're done!
Promise<Cache> getModelCache()	Get Cache object that stores models. This allow for more granular read and writes to the model storage.
EpMode	Enum for endpointer modes

Model object

Function/Object	Description
int findWord(word: string)	Check if a word can be recognized by the model, return the word symbol if word exists inside the model or -1 otherwise. Word symbol 0 is for epsilon
delete()	Delete the model

Recognizer object

Function/Object	Description
acceptWaveform(audioData: Float32Array): string	Processes an audio block and returns a JSON string containing either:

• a partial result ({"partial":"..."})
• or a finalized result ({"text":"..."}) | | setWords(words: bool) | Enables words with times in the output (default: false) | | setPartialWords(partialWords: bool) | Like above return words and confidences in partial results (default: false) | | setNLSML(nlsml: bool) | Set NLSML output (default: false) | | setMaxAlternatives(alts: int) | Configures recognizer to output n-best results (default: 0) | | setGrm(grm: string) | Reconfigures recognizer to use grammar | | setSpkModel(mdl: SpkModel) | Adds speaker model to already initialized recognizer | | setEndpointerMode(mode: EpMode) | Set endpointer scaling factor (default: ANSWER_DEFAULT) | | setEndpointerDelays(tStartMax: float, tEnd: float, tMax: float) | Set endpointer delays | | Promise<void> delete(processCurrent: bool) | Delete the recognizer right after it processes its current block if processCurrent is true; or after processing all remaining blocks if processCurrent is false (default: false) |

Events (legacy / optional)

The partialResult and result events are still available for compatibility, but direct processing of the acceptWaveform() return value is recommended for lower latency and simpler integration.

---

HTTP Remarks

HTTPS

Vosklet is available only in secure contexts (HTTPS)

SharedArrayBuffer

SharedArrayBuffer is necessary to share data between workers, so these response headers must be set:

• Cross-Origin-Embedder-Policy ⟶ require-corp
• Cross-Origin-Opener-Policy ⟶ same-origin

If you can't set them, you may use a hacky workaround in AddCOI.js

Content Security Policy (CSP)

For those who are using CSP, Vosklet's Wasm worker construction will be from a Blob which require the CSP worker-src to include blob:

---

Compilation

• Requires all Autotools commands in PATH, make, and pkg-config. For example, installing with apt would be:

  sudo apt install autotools-dev autoconf libtool make pkg-config

• Changing any option to non-default values requires recompilation

• To remake a specific target, erase its directory in the repo root and run ./make again. Doing this will also remake the final JS

git clone --depth=1 https://github.com/msqr1/Vosklet &&
cd Vosklet/src &&
[Options] ./make
# Example: INITIAL_MEMORY=350mb MAX_THREADS=3 ./make

Option	Description	Default value
INITIAL_MEMORY	Set inital memory, valid suffixes: kb, mb, gb, tb or none (bytes)	315mb as recommended plus a bit of leeway. This memory will grow if usage exceeds this value.
MAX_THREADS	Set the max number of threads (>=1), this should be equal to the number of recognizers used in the program	1
JOBS	Set the number of jobs (threads) when building	$(nproc)
EMSDK	Set EMSDK's path (will install EMSDK in root folder if unset)	../emsdk