class

Whisper::State

Independent inference state for parallel transcription.

Multiple State instances can run inference concurrently on the same model context. Each state maintains its own mel spectrogram, encoder output, and decoder state.

whisper = Whisper.new("/path/to/model.bin")
state = whisper.create_state
# ... use state for inference ...
state.close
whisper.close

Instance methods

Frees the underlying whisper state. Safe to call multiple times.

Source

closed?

Source

decode(tokens : Array(Int32), n_past : Int32, n_threads : Int32 = 4) : Int32

Runs the decoder on the given token context in this state.

Source

detected_language

Returns the detected language ID from the most recent transcription on this state.

Source

encode(offset : Int32 = 0, n_threads : Int32 = 4) : Int32

Runs the encoder on the mel spectrogram in this state.

Source

finalize

Source

lang_auto_detect(offset_ms : Int32 = 0, n_threads : Int32 = 4) : Tuple(String, Hash(String, Float32))

Runs language auto-detection on this state's mel spectrogram.

Source

logits

Returns a Slice of logits from the last decode call on this state.

Source

mel_length

Returns the mel spectrogram length from this state.

Source

pcm_to_mel(samples : Array(Float32), n_threads : Int32 = 4) : Int32

Converts raw PCM audio to mel spectrogram in this state.

Source

set_mel(data : Array(Float32), n_len : Int32, n_mel : Int32 = 80) : Int32

Sets a custom mel spectrogram in this state.

Source

transcribe(samples : Array(Float32), language : String | Nil = "en", n_threads : Int32 = 4, translate : Bool = false, token_timestamps : Bool = false, tdrz_enable : Bool = false) : Array(Segment)

Runs full inference with this state on the given audio samples.

samples must be 32-bit float PCM audio normalized to [-1.0, 1.0], mono, at 16kHz.

Source