Vocametrix Voice Analysis

Extract advanced spectral measures from a sustained vowel: center of gravity, skewness/kurtosis, H1-H2 (breathiness indicator), H1-A1, H1-A3, LTAS slope and tilt, alpha ratio. Returns age/gender-normalized norms and voice pattern classification. BEFORE CALLING: Confirm the user has a sustained vowel recording (/a/ held at comfortable pitch for 3+ s, minimal background noise). If not, explain what it is and ask them to record one. Do not pass connected speech or conversational audio.

Calculate the S/Z phonation ratio (duration of sustained /s/ vs /z/). Normal ratio ≈ 1.0. Ratio > 1.4 suggests vocal fold pathology (the /z/ is shorter). Requires two separate recordings: one of sustained /s/ and one of sustained /z/. BEFORE CALLING: Confirm the user has TWO separate recordings — one of sustained /s/ (voiceless hiss, like a snake) and one of /z/ (voiced buzz, like a bee), each 3–5 s. If not, explain the difference between them and ask the user to record both.

Compute intensity dynamics, pitch-intensity correlation, and composite scores for voice control, projection, stability, effort, and monotonicity. Useful for voice training, public speaking coaching, and vocal fatigue assessment. BEFORE CALLING: Confirm the user has a sustained vowel recording (/a/ held at comfortable pitch for 3+ s, minimal background noise). If not, explain what it is and ask them to record one. Do not pass connected speech or conversational audio.

Calculate the vocal range (ambitus) from a glissando recording. Returns the F0 range (lowest to highest pitch) in Hz, semitones and octaves. It measures pitch only, not intensity, so it is not a full Voice Range Profile (phonetogram). Useful for singers and voice rehabilitation assessment. BEFORE CALLING: Confirm the user has a glissando recording — a continuous pitch sweep from their lowest to highest comfortable pitch on any vowel, without stopping (5–10 s). If not, explain what a glissando is and ask them to record one.

Expert speech therapy assistant providing role-based guidance. Adapts its answers depending on whether the user is a therapist (clinical detail), a patient (accessible explanation), or a parent/caregiver (practical home tips). Maintains conversation context via threadId for multi-turn dialogue.

Analyze text for grammar and syntax errors with severity classification (error/warning/info). Returns overall score, per-issue breakdown, corrected text, and readability statistics. Useful for evaluating written language samples in speech-language assessments.

Classify stuttering disfluency patterns in a speech recording (async, ~30–120 seconds). Returns disfluency types (repetitions, prolongations, blocks), severity score, and fluency rate. The tool polls the result automatically — no separate status call needed. BEFORE CALLING: Confirm the user has a natural connected speech recording (the patient speaking spontaneously or reading aloud). A sustained vowel is not appropriate here — the recording must contain running speech.

Convert French words or phrases to International Phonetic Alphabet (IPA) transcription. Accepts a single string or an array of up to 20 words. Returns IPA transcription per word with optional syllable boundary marks. French only — passing non-French text will return incorrect phonetic output.

Detect French phonemes in an audio recording using a deep-learning classifier. Returns phoneme labels with confidence scores and timestamps. French only. By default uses the baseline model; pass model='logatome-champion' to use the logatome-trained classifier (better accuracy on logatome-type utterances used in speech therapy).

Extract the openSMILE GeMAPS v01b feature set (62 acoustic features, plus speech and silence percentage) from a sustained vowel. Features include F0, jitter, shimmer, HNR, formants, spectral slope and loudness. Commonly used as input to machine-learning voice pathology classifiers. BEFORE CALLING: Confirm the user has a sustained vowel recording (/a/ held at comfortable pitch for 3+ s, minimal background noise). If not, explain what it is and ask them to record one. Do not pass connected speech or conversational audio.

End-to-end therapy plan generation with automatic polling and human-in-the-loop approval. Generates a therapy plan from session data, polls until complete, and presents it for approval. Returns the approved plan or the pending plan awaiting your approval action. After reviewing, call vocametrix_approve_therapy_plan with 'approve', 'modify', or 'reject'. BEFORE CALLING: Confirm that wav2vecOutput comes from vocametrix_classify_stuttering's summary_statistics — do not pass invented or placeholder data.

Run a comprehensive clinical voice assessment in a single call. Executes AVQI, CPP, multi-band HNR, jitter/shimmer, and spectral analysis in parallel, then returns a unified JSON report with all metrics and clinical severity interpretation. AVQI version is derived automatically from the patient language (en/nl/de → v02.03; fr/es/it → v03.01). BEFORE CALLING: (1) Ask for or infer the patient language (en/fr/nl/es/de/it). (2) Show the user the correct connected speech reference sentence for that language (read vocametrix://recording-guide to get it) and ask them to record it. (3) Confirm they al…

Generate personalized speech therapy exercises tailored to patient profile, pathology, and language. Returns structured exercises with instructions, target phonemes, difficulty level, and therapist tips.

Launch an asynchronous LangGraph-powered therapy plan generation from session audio embeddings. Returns a therapy_session_id. Use vocametrix_get_therapy_status to poll progress, then vocametrix_get_therapy_result to retrieve the plan once complete (~30–120 seconds). BEFORE CALLING: Confirm that wav2vecOutput comes from vocametrix_classify_stuttering's summary_statistics — do not pass invented or placeholder data. Passing arbitrary values will produce a clinically meaningless therapy plan.

Generate a word list targeting a specific phoneme with pronunciation hints and difficulty progression. Useful for articulation therapy, phonological awareness drills, and accent training.

Retrieve the completed therapy plan result. Only call when vocametrix_get_therapy_status returns result_available = true or status = 'complete'. Returns the full therapy session with exercise plans, clinical narrative, and HTML report path.

Poll the status of an async therapy plan generation or stuttering classification session. Statuses: pending → processing → pending_approval → complete (or failed). result_available = true means you can call vocametrix_get_therapy_result.

Ingest a publicly fetchable audio URL into Vocametrix cloud storage and return a stable blobUrl. Use this whenever the user provides a direct HTTPS URL pointing to a WAV file — Google Drive direct-download links, S3 public objects, Dropbox links with ?dl=1, signed/presigned URLs, etc. The URL is fetched once and re-stored on Vocametrix; the returned blobUrl is what you pass as audioPath to any analysis tool. This is the right tool when the user's input is a URL. For files attached directly in the conversation (binary content, no URL), use vocametrix_upload_audio with base64 instead.

Interpret a speech-to-text transcription of a spelling attempt and give intelligent feedback. Returns whether the spelling matches, an explanation of differences, and correction guidance. Useful for spelling therapy apps where children spell words aloud.

Translate raw voice metrics (jitter, shimmer, HNR, CPPS, F0, etc.) into clinical-language interpretation with severity classification (normal / mild / moderate / severe) and actionable recommendations. Useful when you have metric values from other tools and want a clinician-readable summary. BEFORE CALLING: Pass metrics that come from Vocametrix measurement tools (e.g. vocametrix_calculate_jitter_shimmer, vocametrix_calculate_hnr, vocametrix_calculate_cpp). Do not invent or estimate values — passing arbitrary numbers produces misleading clinical output.

Measure the sound level over a specified time window in an audio file, in uncalibrated dB (digital full scale is taken as 94 dB, whatever the microphone). Not an absolute SPL: compare values only across recordings made with the same device at the same distance, never between devices or people. Note: startSec must be > 0 (use 0.001 for the start of the file).

Synthesize speech from text using Azure neural text-to-speech. Returns dataUrl, a data URI carrying the whole clip for immediate audio playback. Use dataUrl as the src of an HTML audio element to let the user play the audio. Supports all Azure Neural voice names for the requested locale. BEFORE CALLING: Detect the language of the text. Set speakerLocale to the matching BCP-47 code (e.g. 'es-ES' for Spanish, 'fr-FR' for French, 'de-DE' for German) and set voiceName to a natural neural voice for that locale (e.g. 'es-ES-ElviraNeural', 'fr-FR-DeniseNeural', 'de-DE-KatjaNeural'). Do not leave spea…

Transcribe an audio file using Azure Speech-to-Text with streaming progress. Returns a transcriptionId and streams progress events via SSE until completion. Returns the final transcript and word-level timing. For long recordings, poll the progress events — transcription may take 30–120 seconds.

Upload an audio file the user attached in ChatGPT and obtain a blobUrl. Use this whenever the user attaches a recording in the conversation: pass the attachment as `file`. Then pass the returned blobUrl as the audioPath parameter to any analysis tool.

Upload an audio file to Vocametrix cloud storage and obtain a blobUrl. THIS IS THE MANDATORY FIRST STEP whenever the user attaches an audio file in the conversation — the remote Vocametrix MCP server cannot read your local filesystem or resolve chat-client attachment identifiers. In ChatGPT, use vocametrix_upload_attachment instead: it receives the attached file directly. Workflow: (1) read the attached audio file's binary content, (2) base64-encode it, (3) call this tool with that base64 string, (4) take the returned blobUrl and pass it as the audioPath parameter to any analysis tool (assessm…

Conversational vocabulary tutor adapting to learner profile (native language, target language, age, topic). Uses spaced repetition principles. Maintain conversation context via threadId.