Whisper Transcribe

This component transcribes audio from a video stream.

Component Configuration

Option Description Required

id

A unique identifier for the component instance

displayName

The name shown in the Studio UI of this component

model

The ggml model path

language

Source language (otherwise automatic)

translate

Whether to translate the output to English (default: false)

tinyDiarize

Enable tiny-diarize (speaker-turn detection). Only supported by tdrz models

initialPrompt

Initial prompt to prime the model with expected names, jargon or style. Max ~224 tokens, and prepended to any context from previous chunks

contextPrompt

Prompt each chunk with the tokens recognised in the previous chunk. Off by default, relying only on chunk overlap to resolve partial words

suppressNonSpeechTokens

Suppress non-speech tokens such as [music] and [laughter]

vad

Voice activity detection - use a silero VAD model to skip non-speech audio. Leave unset to disable

stepMs

Audio accumulated before each transcription step (default 3000ms). Lowering this reduces latency but also throughput - if a step is not clearly faster than real time, the workflow will back up

keepMs

Audio kept when clearing the buffer, to allow partial-word recognition (default 400ms)

useGpu

Use GPU if available. This may move load from CPU to GPU rather than increasing throughput (default: true)

numThreads

Number of threads to use. Using a large number of threads rarely improves performance

samplingStrategy

Decoder sampling strategy (default greedy)

maxTokens

Max tokens per segment (0 = no limit)

initialTemperature

The initial decoding temperature, between 0 and 1 (default 0.0)

temperatureIncrement

The increment of temperature on fallback, between 0 and 1 (default 0.2). Set to 0 to disable fallback

noFallback

Do not use temperature fallback while decoding

entropyThreshold

Entropy threshold for decoder fail, similar to OpenAI's compression_ratio_threshold (default 2.4)

logProbThreshold

Log probability threshold for decoder fail (default -1.0)

noSpeechThreshold

No speech threshold (default 0.6)

notes

Notes about this component

Vad

Option Description Required

model

Path to the GGML-format silero-vad model, e.g. from https://huggingface.co/ggml-org/whisper-vad

threshold

Probability threshold to consider as speech (default 0.5)

minSpeechDurationMs

Min duration for a valid speech segment (default 250ms)

minSilenceDurationMs

Min silence duration to consider speech as ended (default 100ms)

maxSpeechDurationS

Max duration of a speech segment before forcing a new segment (default unlimited)

speechPadMs

Padding added before and after speech segments (default 30ms)

samplesOverlap

Overlap in seconds when copying audio samples from a speech segment (default 0.1)

Sampling Strategy (Greedy)

Option Description Required

strategy

"greedy"

bestOf

Number of best candidates to keep

Option Description Required

strategy

"beam_search"

beam_size

Beam size for beam search

Tags: