Vocal biomarkers

How something is said, measured from the waveform.

Pitch, voice quality, timing and articulation from a WAV recording, computed with the same algorithms as the Medera BH-ASR voice layer and checked number for number against it.

For research and clinician review. Not a diagnosis.

medera · /v1/vocal/analyze
sample
Waveform6.0 s · 16 kHz · mono0 s1 s2 s3 s4 s5 s6 spause 0.59 spause 0.48 sspeech100 Hz150 Hz200 Hzmedian 144 HzPitch, every 10 msGlottal cycles · 32 ms6.80 ms6.88 ms6.88 msF0 median144HzPitch variability1.778stJitter (local)0.859%Shimmer (local)8.871%Harmonics-to-noise14.98dBPause ratio0.1981
Explained

Vocal, step by step.

0:59 with on-screen captions. Every number in it is the engine's output for the published sample.

Read the transcript
  1. 01 · How it sounds

    Medera measures how something is said, not what is said.

  2. 02 · The waveform

    Six seconds of speech arrive as a WAV file and are decoded in memory.

    Energy separates speech from silence. Gaps of 250 ms or more are pauses.

  3. 03 · Pitch

    Every 10 ms, autocorrelation finds the fundamental frequency: the pitch.

    The median here is 144 Hz. How much it moves is pitch variability: 1.778 semitones.

  4. 04 · Glottal cycles

    At 32 ms, each cycle of the vocal folds is visible and can be timed.

    Cycle-to-cycle differences in timing and amplitude are jitter and shimmer.

  5. 05 · What comes back

    Every measure follows Praat and eGeMAPS, checked number for number against the Python reference.

    These are the values the API returns for this audio.

  6. 06 · Quality gates

    Every measure has a quality gate: enough voicing, enough cycles, a clean signal.

    If the audio cannot support a value, the API returns null with the reason. It never guesses.

  7. 07 · One request

    One request: send a WAV file and get the vocal panel for each speaker.

    Add a transcript for speaking rate and language; one speaker per channel for response latency.

  8. 08 · Recap

    Measured from the waveform, gated, and compared with the same person over time.

What goes in

Inputs, and only those.

Audio is decoded in memory, measured and discarded when the response is sent. Nothing is stored.

  • 01WAV audio, PCM or IEEE float, 8 to 192 kHz
  • 02Mono, or one speaker per channel for separated speakers and response latency
  • 03An optional transcript, in English or Arabic, for word rates and language markers
How it is measured

Published methods, in order.

Each step names the method behind it. The catalog lists the references for every measure.

  1. 01Find the speech

    Audio is resampled to 16 kHz and speech frames are separated from background with an energy threshold set by Otsu's method.

  2. 02Track the pitch

    Boersma's autocorrelation method with Praat's path finder follows the fundamental frequency every 10 ms.

  3. 03Follow each glottal cycle

    Pulses are tracked by waveform cross-correlation. Jitter, shimmer, harmonics-to-noise ratio and CPPS follow Praat's definitions.

  4. 04Read the spectrum

    Formants by Burg linear prediction; alpha ratio, Hammarberg index and spectral slopes as defined in eGeMAPS.

  5. 05Time the speech

    Pauses of 250 ms or more, speaking and articulation rate from the transcript, and response latency between separated channels.

  6. 06Gate every value

    A value that fails its quality gate is returned as null with the reason, never estimated.

What comes out

The biomarkers, defined.

Every measure links to its catalog entry: definition, method, quality gate, published associations and references.

BiomarkerUnitWhat it measuresPublished associations
Timing
Speaking ratespeakingRateWpmwpmWords per minute over the speaking span, pauses included.RDoC: Sensorimotor systemsDepression↓RPost-traumatic stress↓SSchizophrenia↓MA
Articulation ratearticulationRateWpmwpmWords per minute of speech, pauses excluded.Descriptive
Pause ratiopauseRatioratioShare of the speaking span spent in pauses of 250 ms or more.RDoC: Sensorimotor systemsDepression↑RSchizophrenia↑MA
Mean pausemeanPauseSsMean length of pauses of 250 ms or more.RDoC: Sensorimotor systemsDepression↑RSchizophrenia↑MA
Longest pauselongestPauseSsLongest pause of 250 ms or more.Descriptive
Pause countpauseCountcountNumber of pauses of 250 ms or more within the speaking span.Descriptive
Response latency (median)responseLatencySsMedian gap before this speaker answers the other.RDoC: Cognitive systemsDepression↑S
Loudness peaksloudnessPeaksPerSec1/sPeaks (3 dB or more of prominence) of the smoothed level per second of speech.Descriptive
Voiced segmentsvoicedSegmentsPerSec1/sNumber of continuous voiced stretches per second of speech.Descriptive
Mean voiced segmentmeanVoicedSegmentSsMean length of voiced stretches.Descriptive
Mean unvoiced gapmeanUnvoicedSegmentSsMean length of the gaps between voiced stretches.Descriptive
Voiced fractionvoicedFractionratioVoiced frames as a fraction of speech frames.Descriptive
Pitch
F0 medianf0MedianHzHzMedian fundamental frequency of voiced frames.Descriptive
F0 meanf0MeanSemitonest re 27.5 HzMean F0 of voiced frames on a semitone scale relative to 27.5 Hz.RDoC: Arousal and regulatory systemsAnxiety↑S
Pitch variabilityf0SdSemitonesstStandard deviation of F0 in semitones; low values mean monotone speech.RDoC: Positive valence systemsDepression↓RAnxiety↑SPost-traumatic stress↓S+1
Pitch range (20–80%)f0Range20_80Semitonesst20th to 80th percentile range of F0 in semitones.Descriptive
Voice quality
Jitter (local)jitterLocalPct%Mean absolute difference of consecutive glottal periods over the mean period.Depression~Suicide risk↑S
Shimmer (local)shimmerLocalPct%Mean absolute difference of consecutive pulse amplitudes over the mean amplitude.Depression~Anxiety↑S
Shimmer (local, dB)shimmerLocalDbdBMean absolute dB ratio of consecutive pulse amplitudes.Depression~
Harmonics-to-noise ratiohnrDbdB10·log10(r/(1−r)) at the pitch period, averaged over voiced frames.Descriptive
Smoothed cepstral peak prominencecppsDbdBCepstral peak prominence of the speech after time and quefrency smoothing; a breathiness and dysphonia measure.Descriptive
H1–H2h1h2DbdBLevel of the first harmonic minus the second (uncorrected).Descriptive
H1–A3h1a3DbdBLevel of the first harmonic minus the strongest harmonic near F3 (uncorrected).Descriptive
Loudness and spectrum
LevelintensityMeanDbdBFSMean frame level of speech. Depends on microphone gain, so compare within a setup.Descriptive
Loudness variabilityintensitySdDbdBStandard deviation of the speech frame level.RDoC: Positive valence systemsDepression↓SPost-traumatic stress↓S
Alpha ratioalphaRatioDbdB10·log10 of energy 1–5 kHz over energy 50–1000 Hz in voiced frames.Descriptive
Hammarberg indexhammarbergIndexDbdBStrongest spectral peak 0–2 kHz minus strongest peak 2–5 kHz.Descriptive
Spectral slope 0–500 HzspectralSlope0_500DbPerKHzdB/kHzRegression slope of the dB power spectrum, 0–500 Hz.Suicide risk≠S
Spectral slope 0.5–1.5 kHzspectralSlope500_1500DbPerKHzdB/kHzRegression slope of the dB power spectrum, 500–1500 Hz.Descriptive
Spectral fluxspectralFluxMeanMean squared difference of consecutive normalized spectra in voiced frames.Descriptive
Articulation
F1f1MedianHzHzMedian first formant over voiced frames.Descriptive
F2f2MedianHzHzMedian second formant over voiced frames.Descriptive
F3f3MedianHzHzMedian third formant over voiced frames.Descriptive
F1 bandwidthf1BandwidthHzHzMedian F1 bandwidth.Descriptive
F2 bandwidthf2BandwidthHzHzMedian F2 bandwidth.Descriptive
F3 bandwidthf3BandwidthHzHzMedian F3 bandwidth.Descriptive
F2 spreadf2RobustSdHzHzSpread of F2 over voiced frames, IQR/1.349; a proxy for articulatory range.RDoC: Sensorimotor systemsDepression↓SPost-traumatic stress↓S

Associations are group-level findings from the literature (↑ higher, ↓ lower; MA meta-analysis, R replicated, S single study), never a diagnosis. Compare a person with themselves over time.

Quality gates

Null, never a guess.

When the data cannot support a measure, the value is null and the response says why in quality.flags.

  • Pitch measures need at least 0.5 s of voiced speech.
  • Jitter, shimmer, HNR and CPPS need 1.5 s of voicing, 40 glottal periods, a signal-to-noise ratio of 15 dB and at most 1% clipped samples.
  • Speaking and articulation rate need 5 s of speech and 10 words.
  • Response latency needs one speaker per channel.
POST /v1/vocal/analyze

One request. Try it now.

Each example downloads the published sample and sends it with the public sandbox key. The sandbox takes synthetic data only.

# A six-second synthetic speech sample
curl -s https://www.medera.info/api/v1/samples/vocal -o speech.wav

curl -s https://www.medera.info/api/v1/vocal/analyze \
  -H "Authorization: Bearer mk_sandbox_medera_public" \
  -F audio=@speech.wav \
  -F 'options={"language":"en"}'
POST/v1/vocal/analyze
Sample request

Sends the synthetic sample to the live sandbox with the public key and shows the full response. Nothing personal is sent and nothing is stored.

Response for the sample · 200
{
  "object": "vocal.analysis",
  "api_version": "v1",
  "mode": "sandbox",
  "engine": {
    "name": "medera-inference",
    "version": "1.0.0",
    "vocal": "BH-ASR voice layer 1.2.0 (TypeScript port, number-for-number with the Python reference)"
  },
  "intended_use": "Measured biomarkers and within-person change for research and clinician review. Not a diagnosis and not a medical device.",
  "input": {
    "sample_rate_hz": 16000,
    "channels": 1,
    "duration_s": 6,
    "bits_per_sample": 16,
    "encoding": "pcm"
  },
  "separation": "none",
  "speakers": [
    {
      "label": "speaker",
      "channel": 0,
      "biomarkers": {
        "f0MeanSemitone": 28.58,
        "f0MedianHz": 144,
        "f0SdSemitones": 1.778,
        "f0Range20_80Semitones": 2.807,
        "jitterLocalPct": 0.859,
        "shimmerLocalPct": 8.871,
        "shimmerLocalDb": 0.827,
        "hnrDb": 14.98,
        "cppsDb": 14.3,
        "f1MedianHz": 601.9,
        "f2MedianHz": 1107.4,
        "f3MedianHz": 2361.9,
        "f1BandwidthHz": 86.3,
        "f2BandwidthHz": 89.3,
        "f3BandwidthHz": 227.4,
        "f2RobustSdHz": 678.9,
        "alphaRatioDb": -11.174,
        "hammarbergIndexDb": 28.013,
        "spectralSlope0_500DbPerKHz": 20.852,
        "spectralSlope500_1500DbPerKHz": -22.317,
        "h1h2Db": -3.729,
        "h1a3Db": 10.301,
        "spectralFluxMean": 0.04596,
        "intensityMeanDb": -25.72,
        "intensitySdDb": 8.5,
        "loudnessPeaksPerSec": 5.682,
        "voicedFraction": 0.943,
        "voicedSegmentsPerSec": 4.261,
        "meanVoicedSegmentS": 0.221,
        "meanUnvoicedSegmentS": 0.149,
        "pauseCount": 2,
        "pauseRatio": 0.1981,
        "meanPauseS": 0.535,
        "longestPauseS": 0.59,
        "speakingRateWpm": null,
        "articulationRateWpm": null,
        "responseLatencyS": null,
        "words": null,
        "firstPersonSingularPer100": null,
        "absolutistPer100": null,
        "negativeAffectPer100": null,
        "fillerPer100": null,
        "repetitionPer100": null
      },
      "mfcc": {
        "mean": [
          79.386,
          -7.795,
          -9.465,
          -5.513
        ],
        "sd": [
          21.647,
          12.619,
          20.791,
          12.251
        ]
      },
      "quality": {
        "duration_s": 6,
        "speech_s": 3.52,
        "voiced_s": 3.32,
        "periods": 449,
        "snr_db": 41.3,
        "clipped_fraction": 0,
        "flags": [
          "speakers_not_separated",
          "rates_and_language_require_transcript",
          "latency_requires_separate_channels"
        ]
      }
    }
  ]
}
Computed by the engine for the published sample. Each call also carries an id and a created timestamp. Full reference
FAQ

Vocal, answered.

Still wondering about something?

A real person will answer, usually the same day.

No. It measures the waveform. Attach a transcript to unlock word rates and language markers. Medera BH-ASR, our bilingual speech recognition, provides transcripts, speakers and safety-language review under agreement.

Values depend on the microphone, the room and the task. Compare a person with themselves over time, which is exactly what the inference engine does.

Acoustic measures are language independent. Language markers support English and Arabic.

Combine them

Better together, compared over time.

Send this modality with the others in one call to /v1/analyze, or pass the biomarkers to the inference engine.

Start measuring.

Sandbox today. Production under agreement.

Or make a first request with the sandbox key.