vllm_omni.utils.audio ¶

audio_chunk_pcm_bytes(omni_res: OmniRequestOutput) -> int

Best-effort PCM byte count of the last audio chunk in omni_res.

Used by the audio-streaming continuity tracker to size the player buffer. Returns 0 when the chunk shape can't be interpreted — caller drops the sample rather than recording a wrong byte count.

audio_chunk_sample_rate(omni_res: OmniRequestOutput) -> int

Resolve audio sample rate for the request's audio stream.

mel_filter_bank(
    sr: int,
    n_fft: int,
    n_mels: int,
    fmin: float = 0.0,
    fmax: float | None = None,
) -> Tensor

Compute a mel filterbank matrix.

Drop-in replacement for librosa.filters.mel using torchaudio.functional.melscale_fbanks.

Parameters:

Name	Type	Description	Default
sr	int	Sample rate of the audio.	required
n_fft	int	FFT window size.	required
n_mels	int	Number of mel bands.	required
fmin	float	Minimum frequency (Hz).	0.0
fmax	float | None	Maximum frequency (Hz). Defaults to sr / 2.	None

Returns:

Type	Description
Tensor	Tensor of shape (n_mels, n_fft // 2 + 1).

peak_normalize(
    audio: ndarray, db_level: float = -6.0
) -> ndarray

Normalize audio so peak amplitude reaches a target dB level.

Drop-in replacement for sox.Transformer().norm(db_level=...).

Parameters:

Name	Type	Description	Default
audio	ndarray	Input waveform as a 1-D numpy array.	required
db_level	float	Target peak amplitude in dBFS.	-6.0

Returns:

Type	Description
ndarray	Normalized waveform with the same dtype as audio.