AudioChannelPerformanceConfig¶
class AudioChannelPerformanceConfig ( BaseChannelPerformanceConfig )
Audio channel performance configuration for TTS benchmarking.
Polymorphic Type:
type: audioAll
BaseChannelPerformanceConfigtypes:
Fields:
interactivity_enabledbool =TrueCompute streaming interactivity metrics when timestamp lists are present.
startup_delay_ms_valueslist [ float ] =[0.0, 100.0, 300.0]Fixed delays after first audio to simulate, in milliseconds.
startup_buffer_ms_valueslist [ float ] =[0.0, 100.0, 300.0]Playable-audio targets to simulate before playback, in milliseconds.
min_reportable_stall_msfloat =10.0Playback gaps at or below this duration are treated as transport noise.
fluidity_frame_msfloat =20.0Duration of each complete playable PCM frame used by the Etalon-inspired audio fluidity metric.
fluidity_startup_delay_msfloat =0.0Playback slack after the first complete frame for the primary user_audio_fluidity_index metric. Set this explicitly when the playback client intentionally buffers before starting.
fluidity_attribution_modestr ="conservative"How to attribute fluidity misses to the TTS service. ‘conservative’ emits tts_service_fluidity_index only when all text was sent before the first playable frame. ‘source_oversupplied’ treats every miss as service-caused and must only be used with a trace that guarantees synthesis-eligible text throughout playback. User fluidity is always reported.
persist_raw_timingbool =FalseWrite metrics/audio_raw_timing.jsonl with raw per-event timestamps.
max_expected_audio_msOptional [ float ] =NoneServer-side audio duration cap in milliseconds (e.g. Vajra’s max_decode_tokens * 80ms per token = 163840ms at defaults). When set, any request whose generated audio duration reaches the cap (within one 320ms codec chunk) is counted as suspected_length_cap_truncation in the summary and reported by the run’s health check. The server truncates silently at the cap, so duration-at-cap is the only client-visible signal.