Audio and speech

Dialogue, TTS, voice, cleanup, separation, translation and localization

• articleYouTube multi-language audio (creator-uploaded dub tracks) vs automatic dubbing: eligibility via Advanced features, upload steps, localized thumbnails, copyright-mismatch removal, and YouTube's ">25% of watch time" non-primary-language claim
  Per YouTube Help (R2): multi-language audio lets creators upload their own dubbed tracks (also to Shorts) and, for long-form, language-specific thumbnails; it does not generate tracks. Available to channels with Advanced features. The default track follows the viewer's preferred language (from watch history); translated titles/descriptions are used by search. Upload: Studio (desktop) > Languages > video > Add language > next to "Dub" click Add > audio-only file roughly the same length as the video > Publish. If an automatic dub exists for that language it must be deleted first. Rights warning: if the copyrighted content in a secondary track differs from the original audio, the file may be removed - so music or effects beds must match the original mix. Edits to the original (trim, audio swap) propagate to secondary tracks. Localized thumbnails: upload the main thumbnail, then add per-language thumbnails under Languages. YouTube's own claim: "Creators who uploaded Multi-language audio...
• articleFFmpeg documents r_frame_rate as a guessed least-common-multiple base rate, not the average (mixed 25/30 fps gives 150); avg_frame_rate is the average, and for MOV/MP4 the demuxer accumulates stts durations and frame counts for it (media duration, independent of edit...
  FFmpeg documents r_frame_rate as a guessed least-common-multiple base rate, not the average (mixed 25/30 fps gives 150); avg_frame_rate is the average, and for MOV/MP4 the demuxer accumulates stts durations and frame counts for it (media duration, independent of edit lists).
• articleThe otio-aaf-adapter reader (47886982) contains no volume or gain handling (curator whole-file search: 0 hits), so AAF clip gain is not mapped into OTIO by the...
  The otio-aaf-adapter reader (47886982) contains no volume or gain handling (curator whole-file search: 0 hits), so AAF clip gain is not mapped into OTIO by the reader.
• articleFFmpeg n9.0.2 documents bitexact as writing only platform-, build- and time-independent data, primarily for regression testing; in code it drops the library version from the encoder tag and the local time from the WAV peak chunk, which stabilises bytes but removes version provenance from the...
  FFmpeg n9.0.2 documents bitexact as writing only platform-, build- and time-independent data, primarily for regression testing; in code it drops the library version from the encoder tag and the local time from the WAV peak chunk, which stabilises bytes but removes version provenance from the file.
• articleMediaInfoLib (153ad67a) labels MP4 video VFR when the stts-derived maximum and minimum rates differ by at least 0.001 fps, otherwise CFR; because stts holds decode-time deltas, a file with steady presentation timing but irregular DTS can be reported VFR (consistent with the maintainer statement...
  MediaInfoLib (153ad67a) labels MP4 video VFR when the stts-derived maximum and minimum rates differ by at least 0.001 fps, otherwise CFR; because stts holds decode-time deltas, a file with steady presentation timing but irregular DTS can be reported VFR (consistent with the maintainer statement reported by the producer).
• articleRFC 7845 (Ogg Opus) makes pre-skip a SHOULD-discard count at 48 kHz, while the Opus-in-ISOBMFF draft on opus-codec.org declares the MP4 dOps PreSkip informative only and assigns trimming to the edit list; a tool applying both rules to MP4 could double-trim and one applying neither would keep the...
  RFC 7845 (Ogg Opus) makes pre-skip a SHOULD-discard count at 48 kHz, while the Opus-in-ISOBMFF draft on opus-codec.org declares the MP4 dOps PreSkip informative only and assigns trimming to the edit list; a tool applying both rules to MP4 could double-trim and one applying neither would keep the priming samples (producer synthesis consistent with both texts).
• articleDAWproject at ee4dcdde models clip fades only as fadeInTime/fadeOutTime with a time unit in Project.xsd, although its README feature table lists Crossfades; Project.xsd contains no timecode/SMPTE/frame-rate concept (curator...
  DAWproject at ee4dcdde models clip fades only as fadeInTime/fadeOutTime with a time unit in Project.xsd, although its README feature table lists Crossfades; Project.xsd contains no timecode/SMPTE/frame-rate concept (curator search).
• articlesyncnet_python (907c0b57) conforms video to 25 fps and audio to 16 kHz mono before analysis, reports an integer frame offset (vshift minus argmin of averaged distances) with confidence = median minus minimum, so pipeline offsets are in 25 fps frames (40 ms) regardless of source frame rate; code...
  syncnet_python (907c0b57) conforms video to 25 fps and audio to 16 kHz mono before analysis, reports an integer frame offset (vshift minus argmin of averaged distances) with confidence = median minus minimum, so pipeline offsets are in 25 fps frames (40 ms) regardless of source frame rate; code is MIT while the VGG page offers the model "for research purposes under Creative Commons Attribution License".
• articleWhat numeric dialogue-vs-music/background limits and monitoring methods do the BBC, Channel 4/DPP text and EBU publish for dialogue intelligibility in mixes?
  Corrected 2026-10-23 - see C3-B645-03; the original statements below are kept for history and are superseded where the correction says so. Broadcaster numbers for dialogue clarity are recommendations from different methods and are not interchangeable. The BBC (Sound Mixing guide v5.0.5) suggests pulling music down 4 dB after the final mix and monitoring on a domestic TV set to 65 dB C-weighted with 20 dB pink noise, noting that passing R128 alone does not ensure clarity. Channel 4 DPP v5.2 recommends at least 4 LU dialogue-to-background separation and a factual dialogue LRA of 6 LU or less, as guidance only. EBU R 128 s4 (2023) says the loudness-to-dialogue ratio of cinematic content should not exceed 5 LU (15 LU or more occurs). A Dutch process cited in an EBU review caps voice-to-programme loudness at 4.5 LU.
• articleHow do Dragonframe's audio workspace and Harmony's sound tools break dialogue into per-frame mouth shapes, and what must exist before automatic detection works?
  Dragonframe track reading is manual: you type the dialogue and phonetic cues, Dragonframe splits the words into frame cells, and you drag them into place against the waveform. Face sets come from Photoshop layers, and the reading appears in the X-sheet and can be exported and re-imported. Toon Boom Harmony instead detects phonemes automatically and writes one of eight standard mouth-chart letters per frame, but only if the mouth layer already has drawings named with those letters. The letters are labels for mouth shapes, not the sounds themselves.
• articleHow do ITU-T P.800 and P.808 define the MOS scale, listener rules and crowdsourcing controls, and how do the two Recommendations differ?
  ITU-T P.800 defines the listening-only ACR test whose 5-point mean is the MOS (with a note that older results must be shifted by one to compare) and requires naive listeners (no subjective test in 6 months, no listening-opinion test in a year); P.808 (in force, 2021) adapts ACR/DCR/CCR to crowdsourcing with one gold-standard and one trapping question per 10 stimuli, two-eared headphones and a two-week listening-test gap, and says crowdsourcing complements rather than replaces laboratory tests. Neither defines a TTS-naturalness protocol.
• articleBefore I write an as-recorded dubbing script for Netflix, what must it contain and how must events be timed and labelled?
  As captured in May 2026, Netflix collects only the as-recorded dubbing script (a TTAL transcription of the final dub plus forced narratives), timed so each event runs from the first to the last spoken word, with a new event after pauses of about two seconds or more (reactions and efforts count as pauses), forced-narrative events tagged with on-screen position, due three calendar days before the secondary audio.
• articleWhat archival assets must a dubbing vendor hand Netflix after the dub mix, and how has that deliverables page changed (its change log)?
  As captured on 2026-03-11, Netflix requires dubbing/AD archival assets within 20 business days of final secondary audio via Source Management, split into Standard and Occasional deliverables; an Atmos dubbed dialogue stem must have at least a 7.1.2 bed and at most 118 objects as long-play 48 kHz/24-bit BWAV ADM; the page change log lists 2020-12-02 and 2023-07-13. The live JSON on 2026-10-22 returned only the introduction, so currency after March 2026 is unconfirmed.
• articleBefore I clean dialogue out of a documentary or reality mix, what does Netflix tell nonfiction partners to deliver as M&E, and what has the page changed since 2020?
  For nonfiction, Netflix recommends Foley (a cloth pass gives substantial fill) and, without a Foley budget, married music and effects stems plus a separate dialogue stem from which its dubbing team builds a faux M&E; the Dialogue Guide is the 5.1/5.0 dialogue stem or its mono mixdown, and optional tracks may be Mono to 7.1.2.
• articleWhat belongs in the body of a Netflix M&E, what goes on optional tracks, what must be delivered for M&E QC, and what changed recently?
  Netflix asks for a fully filled, long-play, discrete-channel M&E; language-specific material (non-dialogue vocalisations, language-specific walla and chants, on-screen TV/radio dialogue) goes on separate Optional Tracks rather than the M&E body; a mix reference video is needed only when the stems are not aligned to the latest locked or final cut.
• articleWhat must I label and how must I build beds and objects when delivering a 5.1 or Atmos mix to Netflix as BWAV ADM?
  Netflix's branded BWAV ADM guideline lets beds and objects be any channel width but requires each to carry a Group label of Dialogue, Music, Effects or Narration (from a published list), true-peak limiting at -2.3 dBFS on material routed to the renderer, and 5.1 stems that sum to the printmaster; BWAV ADM delivery is preferred, not required.
• articleWhat changed between ATSC A/85:2013 and A/85:2026-07 for dialogue-gated measurement, target loudness, true peak and streaming?
  ATSC A/85:2026-07 keeps the 2013 exchange targets (-24 LKFS with about +/-2 dB, true peak below -2 dBTP) but adds an approximately +/-0.5 dBTP meter tolerance not present in the 2013 section 6, adds Annex L (streaming) and Annex M (one-page summary), and states that BS.1770-3/-4/-5 share identical loudness and true-peak algorithms; Annex M says long-form dialogue level is measured with BS.1770-1 without the relative gate and describes a -23 to -27 LKFS practical range for streaming operators.
• articleWhich loudness measures should I read (integrated, momentary, short-term, range, true peak) and what do the EBU numbers mean?
  Y1-06's R 128 figures match V5 items h), i), m) and it reports +/-0.5 LU only as the V3 history entry; the C2-B80-AUDIT1 grouping of Y1-06 with a +/-0.5 LU non-live claim is incorrect.
• articleWhich closed-caption (CCAP) properties do libdcp verify and ClairMeta check, which ST 428-10 rules do they miss, and where do their limits come from?
  libdcp (891facc8) cites Bv2.1 (RDD 52) 7.2.6 for both its 3-line and 32-character closed-caption checks; ST 428-10:2008 7.3 caps Text elements per Subtitle at three and mentions 32 characters only in an informative device note; ST 428-10 7.1 forbids overlapping Subtitle time windows.
• articleWhat does YuE's MODEL_LICENSE (YuE2) say clause by clause about hosting, redistribution, modification, training and output use of the weights, and what do THIRD_PARTY_NOTICES add?
  At YuE 1647252d, MODEL_LICENSE licenses the three YuE2 checkpoints under CC BY-NC 4.0 with a dated (2026-09-16) additional permission for individual creators to generate and monetise outputs under responsible-use conditions; it excludes code/tokenizer/evaluation assets from its scope, denies commercial redistribution or sale of weights, and directs companies to a commercial licence. This fills the MODEL_LICENSE gap recorded by L07 for C2-B04-H-12.
• articleWhat are FFmpeg arnndn (RNN speech denoise), adeclick and adeclip for, and which options let you hear what they remove?
  Problem: choose between a model-based denoiser and click/clip repair in an FFmpeg chain. Conditions: FFmpeg filter docs. Method: arnndn reduces noise from speech using recurrent neural networks; the model file option is always required (you must supply a trained model file); mix sets how much processed signal is blended into the output, range -1 to 1, default 1, and negative values keep the removed noise, with -1 letting you hear the actual noise removed. adeclick removes impulsive noise; samples detected as impulsive are replaced with interpolated samples using autoregressive modelling; options include window 10-100 ms (default 55), overlap 50-95 percent (default 75), arorder 0-25 percent (default 2) and threshold 1-100 (default 2; lower detects more samples as clicks). adeclip is the companion for clipped samples. Use (inference): run adeclick before denoise when clicks are sparse, keep threshold at default and lower it only if clicks remain, and use arnndn mix between 0.6 and 1 t...
• articleWhat do Shure's guides say about microphone distance, proximity effect and plosives for voice recording?
  Problem: get a clear, un-muddy voice at the mic. Conditions: single talker, directional (cardioid-type) mic. Method (Shure educational publications and the MV7 guide): keep the mic as close as practical to avoid picking up room reflections and reverberation, but not so close that proximity effect (a rise in low-frequency response with directional mics up close) makes the voice muddy or bassy; aim the mic from below or above rather than straight at the mouth to reduce p/t pops; add an external pop filter, which also helps maintain a consistent distance; keep the mic away from reflective surfaces. Distances quoted: about 6-12 inches from the mouth at roughly a 45-degree angle for a stand or handheld mic in the film/video guide (3-6 inches adds bass emphasis); 1-6 inches for the MV7 podcast mic (more bass, intimate sound, maximum off-axis rejection) or 6-18 inches in the MV7 Far setting. The inverse-square rule: doubling the source-to-mic distance cuts the sound pressure reaching the m...
• articleIn what order and how aggressively should dialogue repair steps be applied, and what do iZotope's authors say about artefact trade-offs?
  Problem: a noisy interview or narration needs several repairs and risks sounding processed. Conditions: iZotope RX workflows described in company blogs (secondary, vendor-authored); AI one-click tools exist too. Method: iZotope's stated goal is a transparent intervention that does not introduce new artifacts that distract the listener. Its noisy-interview list in order: even out levels (Voice leveler), De-clip, Spectral De-noise (learn the background profile), De-ess, then Spectral Recovery for hollow sounding dialogue. A dialogue editor reports putting Mouth De-click first in the chain on a commercial and accepting artifact trade-offs because a consistent music bed hid them. Another editor advises matching ADR to production early, before mixing the original vocal, so processing is done once on a combined track. Adobe's Enhance Speech guide says AI enhancers reduce background noise, remove pops, clicks or dropouts and enhance clarity but that no setup can completely remove unwanted ...
• articleWhat audio does ElevenLabs say Instant Voice Cloning needs (length, quality, format), and why can more audio make the clone worse?
  Problem: make a quick voice clone from a short sample. Conditions: ElevenLabs Instant Voice Cloning (IVC), docs fetched 23 Oct 2026. Method: record about 1 to 2 minutes of clear audio with no reverb, artifacts or background noise; at least 1 minute, and the page says more than 2-3 minutes yields little improvement and can be detrimental to stability. The number of samples is irrelevant; total runtime counts. Quality means how the audio was captured, not the codec, though MP3 at 128 kbps or higher is advised (the same page elsewhere strongly recommends MP3 192 kbps or higher). Why it matters: the AI tries to mimic everything it hears, including speed, inflection, accent, breathing, noise and mouth clicks, so keep tone, performance and level consistent. Example (inference): record a single take reading in the delivery style you want in the final narration, in a quiet treated room, and trim breaths or clicks you do not want copied. Failure modes: clone inherits room reverb or noise; wi...
• articleWhat technical spec does Microsoft give for recording professional custom voice samples (sample rate, level, SNR, noise, compression, silence)?
  Problem: record a voice-talent corpus that a TTS fine-tuning job will accept and that will not bake in noise. Conditions: Azure Speech professional voice fine-tuning (Microsoft Learn, fetched 23 Oct 2026). Method (spec table): WAV mono, 24 kHz, 16-bit PCM (recording at 48 kHz/24-bit is allowed and is down-sampled automatically); peak level -3 dB to -6 dB; SNR above 35 dB; noise before the first word below -70 dB; 100 ms (not more than 200 ms) of silence at the start and end; silence between words/phrases below -30 dB; silence after the last word below -60 dB. Little or no dynamic range compression (maximum 4:1); consistent volume, rate, pitch and style. Audio with SNR below 20 can produce obvious noise in the generated voice; volume lower than -18 dB, waveform overflow (clipping), DC offset or echo and files under 16 kHz are listed as typical errors. Size: usually a reasonable voice from 300 utterances; at least 1,000 (preferably 2,000) for unique-persona voices. Script balance: sta...
• articleHow do you match ADR to production dialogue (EQ, ambience, mic differences) according to iZotope's practitioners?
  Problem: new lines (ADR, or AI-generated dub lines) sound different from the production track. Conditions: iZotope RX Match EQ / Ambience Match workflows, vendor blogs. Method: identify what differs (a different mic, chain and studio; missing outdoor ambience and footsteps). Then (1) Match EQ: bounce a usable bit of the original with the ADR after it, learn the original, render the ADR, then fine-tune by ear (the editor boosted low mids and cut high-mid harshness). (2) Ambience Match: highlight silence containing room tone around the original, learn, and process the ADR; the editor rarely needed more. Do this early, before mixing the original vocal, so one track can be processed together. For missing ambience, Ambience Match applies the ambient sound of the original recording to another clip. Use (inference): for AI-voiced lines, record or generate dry, then match EQ and room tone the same way, and fade a little room tone under edits; check at the real playback level. Failure modes:...
• articleHow does FFmpeg's afftdn denoiser work (noise floor, reduction, profile capture), and what artefact controls does it have?
  Problem: remove steady noise (fan, hiss) from narration in a scripted pipeline. Conditions: FFmpeg with the afftdn filter (FFmpeg filter docs, fetched 23 Oct 2026). Method: afftdn denoises with an FFT. Key options: noise_reduction nr in dB (0.01-97, default 12); noise_floor nf in dB (-80 to -20, default -50); noise_type (white, vinyl, shellac, custom); track_noise tn (adapts the noise floor over time); output_mode (output, noise, input) where 'noise' passes only the removed noise so you can listen to what is being taken out; sample_noise sn start/stop captures a noise profile from a stretch of the input (docs example captures the first 0.4 s: asendcmd=0.0 afftdn sn start,asendcmd=0.4 afftdn sn stop,afftdn=nr=20:nf=-40); gain_smooth gs (0-50, default 0) smooths gains across frequency 'useful to reduce random music noise artefacts'. Docs examples use nr=10, nf=-40. Use (inference): record 1-2 s of room tone at the start of each take for the profile, start with nr near 6-12 dB and rais...
• articleWhat does ElevenLabs Professional Voice Cloning require for training audio (hours, level, distance, noise, single speaker, verification)?
  Problem: prepare a dataset to fine-tune a high-fidelity clone of a narrator. Conditions: ElevenLabs Professional Voice Cloning (PVC) docs fetched 23 Oct 2026; paid plans with PVC slots. Method (docs): use one speaking voice only; style in the samples is replicated, so record in the delivery style you will use (an audiobook voice should be trained on reading); use the language you will mainly need. Amount: 'bare minimum' 30 minutes, 2-3 hours for the best result; total runtime matters, not sample count. Level: roughly -23 dB to -18 dB RMS with a true peak of -3 dB. Mic: any microphone works, but XLR into an audio interface is recommended; distance about two fists from the mic; record in an acoustically treated room (a thick duvet can make a temporary space); no background music or noise; no singing. Processing options in the app can remove background noise or split speakers, but the guidance is to feed clean audio. Verification: verify with the same or similar equipment and a similar...
• articleHow do you use iZotope RX De-reverb (Learn, Reduction, Tail Length) and how do you avoid over-processed dialogue?
  Problem: dialogue recorded in a roomy space sounds distant. Conditions: iZotope RX De-reverb (documentation for RX 6; blog of a later version). Method: Learn on a selection that starts with room tone, is several seconds long and includes the direct signal and reverb tails; it sets Reverb Profile and Tail Length. Reduction is the target wet/dry ratio: larger removes more reverb; negative values add reverb. Reverb Profile adjusts per-band amounts (raise the band where reverb is more prominent). Tail Length approximates RT-60: raise it if tails reappear or early reflections are obvious, lower it if tails and noise floor sound over-processed or dull. Artifact Smoothing controls frequency accuracy; defaults are very high because reverb is smooth in frequency, and lowering it increases artifacts with strong processing. Enhance Dry Signal boosts the direct signal and can help prepare for later de-noising. Output Reverb Only lets you listen to what is being removed. The iZotope blog says pu...
• articleHow should a DIY voice-recording booth for narration or clone training be set up according to Microsoft's custom voice guide?
  Problem: record clean voice without a studio. Conditions: small home setup. Method (Microsoft): a small room with no noticeable echo or room tone, as quiet as possible; drapes on walls to deaden the room; a studio condenser microphone made for voice (a USB mic is simplest because mic, preamp and converter are in one unit; XLR needs preamp and interface); mic on a stand or boom with a pop filter against plosives, ideally with a shock mount; the voice talent stays at a consistent distance (tape marks on the floor if standing; watch chair noise if sitting); a script stand not angled to reflect sound into the mic; engineer in a separate room with talkback; listen to a recording of silence to find noise sources (air vents, fluorescent ballasts, traffic, fans, hum from AC wiring or ground loops) and remove them at the source; set levels to use most of the dynamic range without overdriving; keep the chain simple (mic, preamp, interface, computer). EQ or noise-reduction plug-ins may help, b...
• articleWhat does YouTube's RSS ingestion for podcasts require and not do (ads, verification, private upload first, no back-catalogue notifications)?
  Problem: bringing an audio podcast feed into YouTube. Sources: YouTube Help 'Deliver podcasts using an RSS feed' and 'Publish or disconnect episodes from an RSS feed on YouTube'. Source statements: RSS ingestion is available in select countries/regions; YouTube makes a static-image video from the show art for each chosen episode, uploads new episodes automatically and notifies eligible subscribers. YouTube will not distribute the podcast elsewhere, notify subscribers of back-catalogue additions, auto-update show details or re-uploaded audio, or allow invalid characters such as '<' or HTML in titles/descriptions. Podcast content cannot contain advertisements; branded content (host-read promotions, sponsorships, endorsements) must be declared via the branded-content box. First submission uploads selected episodes as private videos; YouTube recommends waiting until all are uploaded before publishing; check for monetization or copyright issues in the video tab. Steps: Studio, Create, Ne...
• articleOn YouTube, is a podcast a playlist, what makes episodes eligible for podcast features, and what are the stated limits (MP3s, Shorts, ownership)?
  Problem: setting up a podcast show on YouTube without RSS. Source: YouTube Help 'Create a podcast in YouTube Studio'. Source statements: a podcast show is a playlist and episodes are videos in it; it should contain full-length episodes in the order to be consumed; multi-season shows stay in one podcast. Possible features: inclusion in YouTube Music, podcast badges, a spotlight on youtube.com/podcasts, discovery from the watch page, recommendations, improved search. Some playlists are not eligible even if designated podcasts, including content not owned by the creator; copyright-violating episodes may be ineligible; Shorts made to support a podcast do not appear in YouTube Music. To create: Studio, Create, New podcast, or Set an existing playlist as a podcast; add title, description, visibility, square thumbnail. MP3s cannot be turned into podcasts: each episode must be a video. Podcast videos are available for audio-only playback; turning that off requires removing the video from al...
• articleWhat do YouTube's live stream latency options (Normal, Low, Ultra-low) mean, and when should each be used?
  Problem: choosing latency for a live stream. Source: YouTube Help 'Understand live streaming latency'. Source statements: latency is the delay between capture and viewer display; read-ahead buffer is the main source of latency; lower latency means a smaller buffer so viewers feel network problems more and may see more buffering. Normal latency: best for non-interactive streams, highest quality, lowest buffering, all resolutions and live features supported. Low latency: for limited interaction (no waiting for responses, such as polls); most viewers will see under 10 seconds; 4K not supported. Ultra-low latency: for highly interactive streams; most viewers under 5 seconds; may increase buffering. Webcam and mobile streaming are always set up for interactivity and have no latency setting. The setting is under Stream settings in Live Control Room. Method (inference): pick Normal for non-interactive broadcasts, Low for chat answers, Ultra-low for Q&A or live reactions with a stable encod...
• articleWhat usage terms, consent and disclosure duties do the so-vits-svc README and RVC repo state, and how do their licences differ?
  Problem. Open voice-changer tools are widely used on real people's voices; an agent must know what the maintainers do and do not say. Conditions. so-vits-svc 4.1-Stable README; RVC WebUI README and LICENSE. Method. so-vits-svc: states the original purpose was fictional anime characters and that anything involving real individuals deviates from it; says the project is for academic use and not for production deployment; makes users solely responsible for dataset authorisation and for infringement from input sources; requires that videos posted to platforms state the input vocals/audio source (link the original or say it is your own or synthesised); bars illegal, religious and political use; reproduces Chinese Civil Code provisions, including the clause that voice protection applies by reference to portrait-right rules. Licence: AGPL-3.0 (header read). RVC: MIT licence (copyright holders listed); the README text I fetched contains no consent or misuse clause (inference from search; I d...
• articleHow do I check and protect a mix for mono playback (phone speakers, radios) and what causes phase cancellation, according to Sound On Sound?
  Problem. Stereo mixes summed to mono can lose tone or level. Conditions. Any stereo deliverable likely heard on phones, smart speakers or cheap radios. Method (SOS contributors, secondary trade-press). Sum L and R with a mono plug-in or monitor-controller switch; a mono-ising plug-in on the master bus must be removed afterward. Panning a mono source off-centre costs at most about 3 dB in the mono sum, so check that a hard-panned part still functions. Stereo recordings or effect returns with out-of-phase or polarity-inverted content can cancel when summed, often frequency-selectively, so tone suffers; spaced-pair overheads and stereo piano are typical. Use a vectorscope or phase correlation meter (a consistently negative reading indicates a problem); compare against commercial mixes in mono. Failure modes. Polarity-inverted amp simulators, M/S widening. Limits. Articles date 2005 to 2022; music-oriented, dialogue and effects follow the same sum rule (inference). Link to loudness targ...
• articleWhat rights does the NO FAKES Act of 2025 (H.R. 2794 as introduced) create for voice and likeness, and how do licences, minors and union contracts work? (bill, not law)
  Corrected 2026-10-23 - see C3-B614-14; the original statements below are kept for history and are superseded where the correction says so. Problem: understand what a proposed US federal digital-replica right would require of producers and performers. Status: an introduced bill (119th Congress, 9 April 2025); not law as of this text, and the introduced text may have changed. Method (source text): it defines a digital replica as a newly created, computer-generated, highly realistic representation readily identifiable as a person's voice or visual likeness, either in work they did not perform or in a version whose performance was materially altered; it makes the authorisation right a licensable property right that is not assignable during life; post-mortem it lasts 10 years for the right holder plus 5-year renewals on proof of active use, ending no later than 70 years after death. Licence rules: a living adult's licence is valid only up to 10 years, must be written and signed, and must...
• articleWhat does the Seed-VC paper claim about zero-shot voice conversion (timbre leakage, similarity, intelligibility), and how big was the evaluation?
  Problem. Voice conversion must keep what was said and how, while replacing who says it; content features often leak source timbre, and removing leakage can raise word error rate. Conditions. Zero-shot: source and target speakers unseen in training; reference audio of a few seconds. Method. Seed-VC uses a diffusion transformer with in-context conditioning on reference audio and an external timbre shifter during training so content features do not carry the source timbre. Measured result (author-reported). On 100 LibriTTS test-clean source utterances and 8 target-speaker references (seed-tts-eval), Seed-VC reached speaker similarity (SECS) 0.8676 and WER 11.99% against ground-truth WER 8.02% and OpenVoice SECS 0.7547; the paper also reports lower CER than baselines. Failure modes. The authors note OpenVoice scored slightly higher on some DNSMOS quality scores, interpreted as a trade-off with timbre fidelity. Limits. Small eval (8 targets), English-centric, objective metrics only, no l...
• articleVoice changer (speech-to-speech) versus TTS for performance-driven dubbing: what does ElevenLabs say is preserved from the source performance, and what do I have to supply?
  Problem. TTS invents timing and emotion from text; a dubbing or ADR editor who wants a specific performance (whisper, laugh, cry, cadence) needs a method that keeps the actor's delivery and swaps only the voice. Conditions. You have a recorded or live source performance and a target voice (cloned or designed) in the same account. Method. Per ElevenLabs docs, Voice changer takes source audio (file or microphone), a target voice_id and a model (eleven_multilingual_sts_v2 covers 29 languages; eleven_english_sts_v2 is English-only, and the docs say the multilingual model often does better even for English). The docs state the output keeps tone, delivery, cadence, whispers, sighs, laughs and cries; the target voice supplies timbre only. It is positioned as a complement to TTS (fix pronunciation, add performance) and to voice actors (extend range). Failure modes. The docs warn input gain matters: quiet input hinders recognition, loud input clips. There is a remove_background_noise option,...
• articleWhat does Descript document about Regenerate and Smooth jump cuts (inputs, supported footage, batch behaviour, review controls), and what would that mean for AI talking-head clips?
  Problem. Text-based edits leave jump cuts; some tools repair them with AI. Conditions. Descript help pages (23 Oct 2026); English only; uses AI credits on current plans. Method (documented). Regenerate uses AI to smooth or rewrite parts of a video: heal jump cuts, tighten pacing, clean audio, or change what a speaker said (the last needs an authorised custom AI speaker). It currently supports English audio and video only. In video projects it regenerates Audio & video by default, with an Audio only option; sequences, B-roll, multi-speaker recordings and shots where the speaker's face isn't clearly visible cannot have video regenerated; Descript asks consent the first time because the process modifies a speaker's visual performance. Smooth jump cuts scans the whole composition and regenerates audio and video at each jump cut: Find first (lists every cut for review, recommended when the composition has intentional cuts such as scene changes or camera-angle cuts) or Smooth all (5 cuts ...
• articleWhat do Descript's filler-word removal, Edit for Clarity and wordless-media features automate, and what are their documented limits (language, speakers, AI speech)?
  Problem. Text-based editing automates cuts from a transcript; each automation has scope limits. Conditions. Descript help pages 23 Oct 2026. Method (documented). Filler words: detects ums, uhs and similar in English transcripts only, lists each with timestamp and preview, and per instance lets you delete, replace with a gap, ignore, or remove from the transcript. Edit for Clarity: analyses the script to clean pacing, tighten delivery and remove unnecessary or off-topic content; an Edit intensity slider (Low light touch, Medium balanced, Heavy more aggressive); an optional Reword for flow generates AI speech where needed and requires speaker labels and an assigned AI Speaker for each speaker. Wordless media (Beta): shows silence, dead air, ambient noise and pauses as editable script text, disabled by default, enabled per project or in app settings for new projects. Smooth jump cuts can be turned on after these edits to heal the new cuts (P-L1-6-06). Use [inference]: run filler remova...
• articleUnder the NO FAKES bill text, who is liable for unauthorised voice replicas, what are the exclusions, and what takedown and damages mechanics apply? (bill, not law)
  Corrected 2026-10-23 - see C3-B614-14 and C3-B614-15; the original statements below are kept for history and are superseded where the correction says so. Problem: assess exposure for generating or hosting a voice clone. Status: introduced bill, not law. Method (source text): liability attaches to publicly displaying, distributing or making available an unauthorised digital replica, and to distributing a product or service primarily designed to produce unauthorised replicas of an identified person, limited-use for that, or marketed for it. Knowledge: an online service is liable only after receiving a compliant notice or court order (or wilfully avoiding one); others need actual knowledge or wilful avoidance. Exclusions: bona fide news or sports broadcasts, documentary or biographical depiction (not if it falsely suggests authenticity, and not for synchronised music recordings except where First Amendment protected), commentary, criticism, scholarship, satire or parody, fleeting or ne...
• articleIs hitting a mono button on a two-speaker setup enough to judge how a mix sounds on a phone?
  Problem. A mono sum played through two speakers creates a phantom centre image. Conditions. Monitoring mono balance. Method. Hugh Robjohns (SOS) says two-speaker mono sounds slightly over-inflated in bass because both speakers contribute; a single speaker shows the true mono balance, which is what phone and portable listeners hear. Most controllers offer only phantom mono, fine for imaging and phase checks, not bass balance. Failure modes. Forgetting to cancel a mono sum in the output stage yields a mono final mix. Limits. 2005 article; secondary. Version/date. Nov 2005.
• articleWhat are Seed-VC's practical inputs, latency, singing mode and consent-relevant switches, and under what licence is it released?
  Problem. Choosing an open voice converter for a pipeline requires knowing input limits, real-time latency, singing handling and licence. Conditions. Seed-VC main branch README. Method. Reference speech of 1 to 30 seconds clones a voice with no training; optional fine-tuning needs as little as one utterance per speaker (1 to 30 s files, clean, no background music or noise). Real-time mode has stated algorithm delay about 300 ms plus about 100 ms device delay. Singing: set f0-condition to True and use 30 to 50 diffusion steps; auto-f0-adjust is normally off for singing; semi-tone-shift transposes. A v2 option convert-style uses an autoregressive model to convert accent and emotion (false gives timbre-only conversion); anonymization-only ignores the reference and maps speech to an average voice. Failure modes. Noisy or music-laden source or training audio is explicitly undesirable. Limits. Repo README claims, not independently run. Rights. The repo LICENSE is GNU GPL v3 (copyleft; I re...
• articleWhat AI-related changes does the DaVinci Resolve 20.1 guide document for Smart Reframe (pan-only/tilt-only), AI Dialogue Matcher and AI Audio Assistant, and where do they help with AI shots?
  Problem. Prior findings cover Resolve's Magic Mask, Smart Reframe, Voice Isolation and SmartSwitch; the 20.1 release adds controls worth knowing. Conditions. Blackmagic DaVinci Resolve 20.1 New Features Guide (August 2025); the extracted pages do not mark these items as Studio-only (the guide marks Apple Spatial Audio Format as Studio-only), so edition availability is unverified. Method (documented). Smart Reframe gets a Mode dropdown to limit AI reframing to a specific axis: Auto adjusts pan and tilt as it deems necessary, Pan Only adjusts left-right only, Tilt Only up-down only. AI Dialogue Matcher: applying a Dialogue Matcher profile to a timeline clip adds an instance in the Inspector's Effects tab with controls for the amount of Voice Isolation and a Dry/Wet mix to match the ambient quality of the source more accurately. AI Audio Assistant: option to switch between the macro view and the full view of any plugin applied during mixing; disabled tracks are ignored during mixing; q...
• articleFor singing voice conversion, how do Seed-VC and so-vits-svc handle pitch, and what did Seed-VC measure against RVCv2?
  Problem. Speech VC settings damage singing: automatic pitch prediction drifts and pitch range mismatches produce wrong notes. Conditions. Converting vocal stems (isolated, dry) to another timbre. Method. Seed-VC conditions the model on source F0 (extracted with RMVPE, quantised into 256 bins) and shifts F0 by gender or semitone setting between source and target; its README says to enable f0-condition and use 30 to 50 diffusion steps for singing. so-vits-svc warns not to use auto_predict_f0 for singing because it can cause serious pitch issues, and says silence-slicing minimum interval can be lowered to about 100 or 50 for singing. Measured result (author-reported, Seed-VC paper Table 2): on singing, Seed-VC SECS 0.7405 vs RVCv2 0.7264, both F0 correlation about 0.94, Seed-VC lower CER, RVCv2 slightly better DNSMOS. Failure modes. Wrong octave or key shift; auto pitch on; accompaniment bleed in the input (RVC README points to Ultimate Vocal Remover for stem separation; inference that...
• articleIf I run a non-native or differently accented actor through ElevenLabs Voice changer, will the target voice's accent replace theirs?
  Problem. Producers assume a voice swap also swaps the accent or language character. Conditions. ElevenLabs Voice changer with any target voice. Method. The docs state the input determines accent and cadence: a British target voice fed an American-accented recording yields that voice with an American accent; a Portuguese-accented input keeps the accent. So, to get a given accent, cast or direct the source performer for it (or use TTS/dubbing for accent change). Failure modes. Treating the library voice's label (British, American) as a guarantee of accent. Limits. Vendor best-practice statement only; no measurement of accent fidelity. The Seed-VC repo, by contrast, exposes a convert-style switch that attempts accent and emotion conversion with an autoregressive model (see related finding), showing that accent retention is a design choice that differs by tool. Version/date. Docs as fetched.
• articleWhere does AI video-to-audio fall short of a Foley pass on surface and timbre control?
  Problem. A Foley artist changes shoes, surfaces and props per scene; V2A gives what the model infers. Conditions. Current open V2A models (MMAudio-class) with text or video conditioning. Method. AC-Foley (ICLR 2026) states existing systems cannot synthesise creator-specified variations such as footsteps varying by surface material, attributing it to coarse training labels and text that cannot describe micro-acoustic features; its remedy is conditioning on reference audio. Its own limitations section adds that with several concurrent sources (dialogue, ambience, object interactions) the model may fail to align specific sounds with their visual triggers. Use: for hero footsteps and cloth, treat V2A as a draft layer; supply reference audio of the surface or record the pass. Failure modes: same footstep timbre on all surfaces; ambiguous mixed scenes. Limitations. Authors' own framing of related work; results are their benchmarks, not a human-Foley comparison; none of this was run by me....
• articleIs the 0.4 s rectangular Momentary window universal? Correction to C3-B539-01 (missing limitation): it is the 'EBU Mode' definition; ITU-R BS.1771-1 prescribes a first-order IIR filter with a 0.4 s time constant, so meters can differ on Momentary readings.
  Problem. C3-B539-01 states Momentary loudness uses a sliding rectangular 0.4 s window as the Tech 3341 definition, without noting that this is specific to EBU Mode. Conditions. Choosing or comparing loudness meters. Method. Tech 3341 says further attack/release slowing after the sliding rectangular windows is not allowed in EBU Mode, and notes that ITU-R BS.1771-1 prescribes a first-order IIR filter with a 0.4 s time constant to measure Momentary Loudness; EBU kept the rectangular-window definition in EBU Mode until evidence clarifies the issue. So a meter following BS.1771 Momentary ballistics may read differently from an EBU Mode meter. Short-term (3 s) and Integrated gating statements in C3-B539-01 are unaffected. Failure modes. Treating any 'Momentary' reading as an EBU Mode 0.4 s rectangular window. Limitations. The IIR detail is taken from the bracketed note in V4 (Nov 2023); later ITU revisions were not checked. Version and date. EBU Tech 3341 V4, read 23 Oct 2026.
• articleDid Illinois HB4875 become law, and when does it take effect?
  Problem. The earlier note on enrolled HB4875 left enactment and effective date unverified. Method. The Illinois General Assembly bill-status file for HB4875 (103rd GA) shows Governor Approved on 8/9/2024, Effective Date January 1, 2025, and Public Act 103-0836. So the digital-replica amendments to the Right of Publicity Act are in force; the enrolled text previously summarised applies subject to any later amendment I did not check. Use: treat Illinois consent and platform-liability rules as live for uses in Illinois from 1 January 2025. Failure modes: assuming a bill status page equals text in force; later 104th GA amendments were not checked (the HB4875 number in the 104th GA is an unrelated procurement bill). Limitations. Status file only; I did not re-read the Public Act text or the statute as codified. Version/date. Fetched 2026-10-23. Ledger: Tennessee HB2091 (ELVIS Act) page shows signed 03/21/2024, Public Chapter 588, effective 07/01/2024, matching earlier items.
• articleProblem. C3-B506-01 says return_speech 'also returns separated dialogue as its own file' and recommends mixing in an editor with dialogue kept via preserve/keep-speech, without the limitation the schema states. Conditions. Mirelo SFX 2.0 video-to-audio on fal with preserve_speech true (the...
  Problem. C3-B506-01 says return_speech 'also returns separated dialogue as its own file' and recommends mixing in an editor with dialogue kept via preserve/keep-speech, without the limitation the schema states. Conditions. Mirelo SFX 2.0 video-to-audio on fal with preserve_speech true (the default). Method. The schema says that with preserve_speech true the returned file is the speech separated from the input around which the soundtracks were generated, not the speech in the soundtracks; the model renders that speech anew, so the file does not cancel or double it cleanly when layered onto a soundtrack. With preserve_speech false the soundtracks are effects only and speech is separated while they generate. The file is mono, in the chosen upload format, as long as the soundtracks, and not billed. Failure modes. Layering the stem over a preserve_speech soundtrack to rebuild or null dialogue can double or comb the voice. Limitations. Behaviour read from the schema description only; no a...

More