Voice Match — same-voice check & TTS consistency (speaker similarity)

Ask a simple question over an API: "is this the same voice as my reference?" Give a reference audio and one or more candidate audios, and get back a similarity score (0–1) plus a same / different-voice decision for each — ranked, with the best match flagged. Built especially for TTS consistency QC: when you generate a voice several times and it drifts, run your takes through this and keep the ones that match (or regenerate until one passes). Runs on CPU, no GPU, no API key.

• 🎯 Same voice? — reference vs candidate → cosine similarity + pass/fail
• 🔁 Batch / ranking — one reference vs many candidates → sorted by similarity, best match flagged (perfect for "generate N takes, pick the matching one")
• 🎚️ Selectable strictness — tts (same synthetic voice, strict), strict, balanced, tolerant (noisy/phone), or a custom threshold
• 🧠 ECAPA-TDNN speaker embeddings (near state-of-the-art), with ffmpeg preprocessing (16 kHz mono, silence-trim, loudness-normalize) for accuracy
• 🛡️ AI-voice flag (advisory) — every audio also gets an aiGenerated signal (is it a deepfake/synthetic voice?). Keep it on for QC, or run operation: "detect" to only screen audios for AI generation, no reference needed.
• 💵 Pay per comparison · no key · no GPU

The use case it was built for

You generate TTS (Google, ElevenLabs, our own AI Voice Generator…) and across takes the voice varies — you end up regenerating until it "sounds right". Instead: pick one good take as the reference, generate a batch, send them here, and the Actor tells you which takes match the reference and by how much. Automate the loop: accept the first candidate that passes, or regenerate if none do.

Sample output (batch)

{
  "ok": true, "level": "tts", "threshold": 0.72,
  "reference": { "durationSec": 6.4 },
  "count": 3, "matched": 2,
  "best": { "index": 1, "score": 0.94, "sameVoice": true },
  "ranked": [
    { "index": 1, "score": 0.94, "sameVoice": true,  "durationSec": 5.9 },
    { "index": 0, "score": 0.78, "sameVoice": true,  "durationSec": 6.1 },
    { "index": 2, "score": 0.55, "sameVoice": false, "durationSec": 6.0 }
  ]
}

Input

• referenceAudio — the voice to match against (http(s) URL or base64). ~5–10 s of clean speech works best.
• candidateAudios — one or more audios to compare (your several takes).
• level — tts (default, strict; same synthetic voice), strict, balanced, tolerant (noisy). Or set threshold (0–100, e.g. 70 = cosine 0.70) to override.

AI-voice detection (advisory)

Alongside the similarity check, each audio is scored by a deepfake-audio classifier (wav2vec2) and gets:

"aiGenerated": { "likelyAiGenerated": false, "aiConfidence": 0.03 }

Leave detectAi on to flag synthetic/cloned audio during QC, or set operation: "detect" and pass audios to screen a batch with no reference. This is an advisory signal, not a verdict — modern TTS can be very clean and detection is probabilistic; treat it as a heads-up, not proof.

How similarity maps to a decision

The score is the cosine similarity of the two voices' ECAPA-TDNN embeddings. Same-voice pairs score high (typically 0.7–0.95, and near-identical TTS takes even higher); different voices score lower. The chosen level sets the cutoff for sameVoice. You always get the raw score too, so you can tune the cutoff to your data.

Pricing

Per comparison from $0.02 (drops to $0.008 on higher tiers). A batch of 5 candidates = 5 comparisons. Platform usage included. No GPU cost — this runs on CPU.

Related Actors

• AI Voice Generator & Cloner — generate/clone the voices, then verify consistency here.
• ffmpeg API — convert/trim audio before comparing.

Accuracy & honest limits

This is voice biometrics, which is probabilistic, not a guarantee — the score is a confidence, not proof of identity. Accuracy is high with clean, ≥5 s clips and drops with noise, very short audio, heavy compression, or cross-condition recordings (studio vs phone). For your TTS-consistency use case (same generator, clean output) separation is large and results are very reliable. The aiGenerated flag is an advisory deepfake signal, not a guarantee — a good AI clone can still score high against the original by design, so this is not a standalone security/anti-fraud authenticator (that needs liveness + full anti-spoofing).

Privacy

Voice is sensitive personal data. Only compare audio you have the right to use, and don't use this to identify or track people without their consent.