Bezahlte Endpunkte für autonome Agenten. Abrechnung über x402 in USDC auf Base Mainnet.
POST /voice-evidence
Returns measured evidence about a recording, not a verdict on it. The spectral cut-off says which lossy stages the file has already been through, since every encoder discards everything above its own ceiling. The room tone measurement compares the noise floor across the silent stretches: a microphone in a room produces a floor that differs between pauses, and a level identical to the hundredth of a decibel across every pause means the silence was inserted rather than recorded. Loudness range, true peak, clipped sample runs and DC offset describe the processing chain. Container and stream tags are reported verbatim, because an encoder tag contradicting the claimed origin is evidence that needs no interpretation. What this endpoint deliberately does not return is a judgement on whether a voice is human or synthetic. That judgement requires a classifier calibrated against synthetic speech, and generating or holding synthetic voices is forbidden in this hub in use, integration and simulation alike; a verdict without that calibration would be a number whose error nobody has measured. Every marker therefore comes with its method and its limits, so the caller can combine it with evidence this service cannot see.
0.006 USDC je Aufruf · MCP audio_voice_evidence
POST /speech-rate
Transcribes the recording, then measures rate rather than words. Two figures come back and they are not interchangeable: the gross rate over the whole running time, which is what a listener experiences, and the articulation rate over speaking time only, which is what a speaker can actually change. On a recording with normal pauses the two differ by thirty per cent or more, and coaching against the wrong one produces the wrong correction. The rate is also reported per segment with the fastest and slowest passages named by timestamp, so a passage that runs away is findable rather than merely counted. Pauses are measured independently of the transcript, by level detection, because a transcriber places segment boundaries by grammar and not by silence. The comparison scale for the verdict is stated in the answer and is drawn from broadcast practice: roughly 120 to 150 words per minute for narration, 150 to 180 for conversational presentation, above 200 for auctioneer-grade delivery. Language is detected, not assumed, and reported, since the same rate means different things across languages. This endpoint sells the measurement, not the transcript; for a transcript with per-word confidence, medienwerk /v1/stt is the endpoint for that.
0.008 USDC je Aufruf · MCP audio_speech_rate
POST /duck
Takes a voice recording and a music bed and returns one mixed file in which the music is pulled down by the voice through an ffmpeg sidechain compressor. Threshold, ratio, attack and release are all settable and all reported back with the result, because they are the difference between a mix that breathes and one that pumps: a release under 100 ms makes the bed audibly flutter between words, above 1500 ms it never returns between sentences. The bed is looped or cut to the length of the voice track, and which of the two happened is stated. After mixing, the result is measured rather than assumed: the integrated loudness of the passages where the voice speaks is compared against the passages where it does not, and the difference is the ducking depth that was actually reached — the figure a producer checks and the one a parameter set alone cannot promise. Both input tracks and the mix are reported with their loudness in LUFS and their true peak, so the mix can be dropped into a chain with a known target. The result comes back inline as base64; at six megabytes the answer is refused with the reason rather than truncated.
0.010 USDC je Aufruf · MCP audio_duck
POST /trim-silence
Finds silent stretches with ffmpeg level detection and shortens each one that exceeds the threshold length. Pauses are shortened to a floor, never removed outright: a pause cut to zero makes the end of one sentence collide with the start of the next, which is a worse defect than the dead air it replaced. The floor defaults to 400 milliseconds, roughly the pause a speaker leaves between sentences, and is settable. Leading and trailing silence are treated separately and can be removed completely, because there the collision problem does not arise. Every cut comes back with its start and end in the original timeline, its original length and its new length, so the edit is auditable rather than a black box — with a long recording that list is the difference between accepting the result and having to listen to the whole thing. The saving is reported in seconds and as a share, and the loudness of source and result are both measured, since trimming pauses raises the integrated loudness of the file and the next stage in a chain needs to know by how much. The result comes back inline as base64.
0.008 USDC je Aufruf · MCP audio_trim_silence
POST /loudness-normalise
Normalises to −14 LUFS by default, the level the streaming platforms converged on, with named presets for podcast, European and United States broadcast, or a free target. The correction runs in two passes: the first measures integrated loudness, true peak, loudness range and threshold, the second applies a constant correction derived from those figures. The single-pass form of the same filter works its correction out as it plays and therefore rides the level during the opening seconds of a file — audible on every track, and the reason the measurement pass is not optional here. The measured values from both passes come back in full, including the loudness range, because a recording with a range above roughly 15 LU cannot be brought to a target without either compressing it or clipping the peaks, and which of the two happened is something the buyer has to be able to see. True peak is limited to the target ceiling so that a lossy encoder downstream does not push the signal into distortion — normalising to a sample peak instead of a true peak is the most common way a file that measured clean arrives clipped. If the recording already sits within half a decibel of the target, that is stated and the correction applied is near zero rather than being dressed up as work.
0.008 USDC je Aufruf · MCP audio_loudness_normalise
POST /speech-duration
Counts syllables rather than words, because word length differs between languages far more than syllable duration does — a German compound and its four-word English equivalent take about the same time to say and would give wildly different word counts. On top of the syllable count come three corrections that a plain rate multiplication misses: digits are counted at the length they are spoken rather than written, since a year is four characters and five syllables; letter-by-letter acronyms are counted the same way; and punctuation adds pause time, with a sentence break worth more than a comma. Four delivery rates are offered from deliberate reading to advertising copy, each with the kind of material it describes and its syllables per second stated, and a free rate can be given instead. The answer carries the syllable count, the pause time, the rate used and the resulting seconds separately, so an estimate that comes out wrong against a real recording can be corrected by changing one number instead of discarding the whole result. A range is returned alongside the point estimate, because no speaker holds a constant rate and a single number pretends to a precision this method does not have. No audio is involved and no speech is generated: this is arithmetic over text, and it is offered for planning a recording session, not for replacing one.
0.002 USDC je Aufruf · MCP audio_speech_duration
POST /voice-spectrum
Computes the power spectrum from raw samples at 128 logarithmically spaced probe frequencies and sums it into six bands whose boundaries are fixed and stated in every answer: rumble below 80 Hz, the fundamental range to 250 Hz, body to 800 Hz, the second formant range to 2 kHz, the consonant range to 6 kHz, and air above that. Naming the boundaries is not decoration — a bass share quoted without them cannot be compared with anyone else's, and most tools do not say where theirs run. The answer reads the distribution against two practical questions. Whether the recording survives a small speaker is decided by the share below 250 Hz that a phone or laptop speaker simply cannot reproduce, so a voice whose energy sits mostly there arrives thin. Whether it is intelligible is decided by the 2 to 6 kHz band that carries the consonants, and a share too low there is heard as mumbling regardless of how loud the recording is. The fundamental of the voice is estimated over the whole recording and reported with what it does not survive: more than one speaker, and a microphone with a low-frequency roll-off. The spectral cut-off is reported alongside, because it says which lossy stages the file has already passed through and therefore which of the missing air was never there to begin with.
0.006 USDC je Aufruf · MCP audio_voice_spectrum
POST /convert
Converts to FLAC, WAV at 16 or 24 bit, MP3, Opus or AAC, with sample rate, channel count and quality all settable and all reported back. What separates this from a bare transcode is the verdict on what the conversion actually did. The spectral cut-off of the source is measured first, so a source that already passed through a lossy stage is recognised before it is wrapped in a lossless container: converting an MP3 to FLAC multiplies the file size by roughly ten and restores nothing, and the answer says so with the measured cut-off as the evidence rather than leaving an archive to fill up in good faith. A conversion that discards information the source carried is named as such, with the specific loss identified — a lossy target, a lower sample rate, a downmix to mono or a reduction in bit depth are four different losses and are reported separately. Loudness is measured before and after, because a lossy encoder shifts true peak upward and a file that measured clean can arrive clipped. Downmixing to mono is offered and flagged, since for a single voice it halves the size at no cost and for anything stereo it is destructive. The result comes back inline as base64; a result above six megabytes is refused with the reason and the size it would have had.
0.006 USDC je Aufruf · MCP audio_convert
POST /ad-slot-auction
Four actions on one endpoint. create opens a slot with its position, length, reserve price and closing time, and returns an admin token that only the opener holds. bid places a sealed bid, checked against the reserve, the slot length and the publisher's excluded categories before it is accepted — a rejected bid is recorded with its reason rather than silently dropped, so a bidder can see why it never competed. close settles the auction and needs the admin token: the highest valid bid wins and pays the second-highest price, or the reserve if it was the only bid. status reads the auction, and before the close it returns the number of bids but never their amounts, to anyone, the publisher included. The second-price rule is the point of the design rather than a detail: under it a bidder's best move is to bid what the slot is genuinely worth to them, so the publisher learns real values instead of guesses about the competition. What this is not is stated in every answer: it is a ledger and a clearing mechanism, it holds no funds, it delivers no audio, and the winner still owes the publisher out of band. Every state change is recorded with a timestamp so an auction can be audited after the fact.
0.004 USDC je Aufruf · MCP audio_ad_slot_auction