Audio and speech

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

• articleHow do speech-to-3D-face tools relate to 2D lip-sync tools and ARKit blendshapes, and when should I choose each?
  Problem. Choosing between 2D video lip-sync and audio-driven 3D facial animation. Conditions. A character that already exists as a rig or can be built as one. Method (inference). 2D lip-sync tools in C3-B576-06 repaint the mouth region of existing video, so identity and lighting come from the footage; 3D tools such as C3-B790-02 output ARKit-style weights or vertices, so identity, lighting and camera are yours to render and the animation can be edited, retimed and reused. Pick 3D when the character is a rig, needs editable curves, or must be rendered from several angles; pick 2D when the footage is fixed and only the mouth must change. The ARKit 52 names are the shared interface, as in C3-B752-08. Failure modes. Using a 3D route without a layer for eyes, brows and idle motion (see C3-B790-03). Limitations. This synthesis is mine and untested by me; sources say nothing on a head-to-head comparison.
• articleWhat speech languages and emotions was Audio2Face-3D trained on, and how does emotion control work?
  Problem. Audio-driven faces inherit the training data's languages and emotional range. Conditions. Choosing a model for a non-English or highly emotional character. Method. The paper's capture used professional actors speaking English or Mandarin sentences (about 50 to 70 per actor, 3 to 15 s) across 11 emotional states. Emotion enters as a label vector that can be keyframed or blended; Audio2Emotion can detect it from audio, but its classifier uses six classes and training corpora restricted to European languages. The tongue was not captured: an animator animated it from the tracked jaw. Failure modes. Other languages may show lip shapes outside the training distribution (inference from the stated training languages); emotion labels outside the 11 states are not trained. Limitations. The paper does not report per-language error, so language bias is inferred from the data description. Version/date. arXiv 2508.16401; model v3.0.
• articleWhat are Audio2Face-3D's documented failure cases (noise, non-speech sounds, upper face, idle and listening motion)?
  Problem. Speech-driven faces look right on clean dialogue and wrong elsewhere. Conditions. Applies to Audio2Face-3D as described by its authors. Method. Prepare audio accordingly: use clean, dry speech stems, hand-animate or layer non-verbal sounds (breaths, laughs), and plan separate eye, brow and idle layers. The authors state that severe noise, strong non-verbal sounds, or lip shapes outside the training data can break lip accuracy; upper face and eyeball motion is simple because the model sees only a short audio window and lacks semantic context; and it cannot generate natural idle or listening motion for long silences or a conversation partner's speech. Failure modes. A conversational avatar that only listens will look frozen unless idle or listening motion is added elsewhere (inference). Limitations. Limits are the authors' own statement, with no numeric thresholds. Version/date. arXiv 2508.16401 (2025).
• articleWhich parts of NVIDIA's open-sourced Audio2Face-3D stack are under which licence, and can I use the models and sample data commercially?
  Problem. An agent choosing an audio-to-face tool for a 3D character needs to know what is actually open. Conditions. Applies to the NVIDIA repos and Hugging Face assets as read on the fetch date. Method. Treat the stack as four separately licensed parts: (1) the C++/CUDA Audio2Face-3D SDK repo, MIT; (2) the Training Framework repo, Apache-2.0; (3) the pretrained model weights on Hugging Face (Audio2Face-3D-v3.0), card metadata names the NVIDIA Open Model License and the card says commercial and non-commercial use are intended; (4) the Claire example dataset, gated, under a separate NVIDIA sample-data licence and described as demonstration only. Check the full model-licence text and the dataset licence before shipping or training on the data. Failure modes. Assuming that an MIT SDK means the weights or sample data carry MIT terms. Limitations. I read licence names and short statements only; I did not read the NVIDIA Open Model License text or the dataset LICENSE body, and nothing her...
• articleWhat does Riverside say Magic Audio does, and how should I treat the claim next to my own cleanup chain?
  Problem. Deciding whether to rely on built-in AI enhancement for remote recordings. Conditions. Material recorded in Riverside. Method. Riverside describes Magic Audio as an AI enhancer and equalizer that removes background noise and emulates a richer sound, regardless of microphone. Failure modes. Enhancement applied destructively would remove the raw option; Riverside's FAQ says edits are non-destructive (not tested). Limitations. The page contains an on/off demo, not a measurement; nothing here is a result. Version and date. Fetched 2026-10-23.
• articleWhat does ipDTL say it records and uses for remote interviews and voice sessions?
  Problem. Voice-over and interview sessions need a live link plus a local-quality record. Conditions. Browser-based sessions with a link. Method. ipDTL's home page lists a virtual mixer with multitrack WAV recording and a wideband Opus codec, with up to 6 remote connections and a use case labelled Voice Sessions. Failure modes. Opus is lossy on the wire, so the multitrack recording may be of the received stream rather than a local lossless file (inference; the page does not say). Limitations. No sample rate or file naming convention was readable. Version and date. Fetched 2026-10-23.
• articleWhat do remote-recording vendors say about sync and drift when you do a manual double-ender?
  Problem. When each person records locally with their own software and files are exchanged, tracks must be aligned by hand. Conditions. No platform handling upload and alignment. Method. Riverside's blog (a competing vendor with an interest) describes double-ender as best quality but manual and error-prone, and says platform recording removes sync issues. A common mitigation (a shared clap or tone at start) is my own inference, not in this source. Failure modes. Sample-clock differences between machines can cause slow drift over long takes (inference, not documented in the fetched page). Limitations. No page fetched gives drift numbers; no measurement made. Version and date. Fetched 2026-10-23.
• articleWhat does Zencastr say about when each participant's local track reaches the cloud and how many tracks I get?
  Problem. Knowing when local tracks are safe in the cloud. Conditions. Zencastr hosted session. Method. Support says each participant is on a separate track and tracks are downloadable individually or as a ZIP from the Library. Zencastr's blog says each track is captured locally and pushed to the cloud when the session ends, which differs from Riverside's upload-while-recording description. Failure modes. A participant closing the page or losing power before the push completes (inference from the blog's design; the blog says a failing laptop leaves the WAV safe, a claim not tested here). Limitations. Blog wording may be outdated against the current product. Version and date. Fetched 2026-10-23.
• articleWhat audio format do Riverside and Zencastr actually state for local audio tracks (sample rate, bit depth, container)?
  Problem. Remote-recording tools differ on stated format and the docs do not always agree. Conditions. Choosing a platform for interview or voice-over capture. Method. Zencastr's support page states audio as 48 kHz WAV or MP3, one track per participant. Riverside's blog states 48 kHz, 16-bit (1,536 kbps) for audio, and its local-recording blog says uncompressed 48 kHz WAV. Zencastr's own blog advises 48 kHz 24-bit capture, but that is a recommendation, not a statement of what the tool records. Failure modes. Assuming 24-bit from a recommendation; assuming one default across plans. Limitations. Bit depth for Zencastr is not stated on the support page read. Version and date. Fetched 2026-10-23.
• articleHow does local (double-ender style) recording in Riverside work, and why does it survive a bad connection?
  Problem. Remote interviews recorded only as a call stream inherit the connection's quality. Conditions. A browser-based remote studio where each participant records on their own device. Method. Per Riverside's FAQ each participant's audio is captured on their own device and uploaded to the cloud in the background while the session runs; its blog calls this progressive uploading (small chunks). Up to 10 participants (9 guests plus host) get separate tracks. Failure modes. Riverside's blog lists crashes/disconnects as the risk of cloud-only recording, which progressive upload is meant to cover; this is a vendor statement not a test. Limitations. Nothing here measures how much audio is lost after a crash. Version and date. Pages fetched 2026-10-23 undated.
• articleWhy do theatrical films get turned down by AC-3 dialogue normalisation at home, and what dialogue levels does Katz give?
  Problem. Theatrical dialogue sits louder than the -31 dialnorm target, so a film carried to AC-3 home delivery is reduced. Conditions. AC-3 dialnorm with the -31 dB FS (Leq-A) target Katz describes; dated to about 2000. Method. Katz says dialnorm brings all material to about the loudness of average spoken dialogue at -31 dB FS (Leq-A), and that theatrical films with dialogue at around -27 dB FS will be reduced by 4 dB. Example. Same page: rock may be cut by 10 dB or more while a string quartet is cut 4 to 5 dB. Failure modes. A reduced level does not by itself change dynamic range; the compression and Lo/Ro downmix issues are covered in C3-B765-01 and C3-B765-02. Limitations. The page expresses -31 as Leq-A and -27 as dB FS without detail; modern streaming uses integrated loudness (LKFS) targets from the platform spec, which I could not read. Version/date. Article updated from September 2000.
• articleHow does a theatrical mix differ in monitor level and dynamic range from a home/near-field mix, according to Bob Katz?
  Problem. Mixing a film for cinema and re-using it for home produces complaints about dialogue and loud effects. Conditions. Katz's article, updated from the AES Journal September 2000, concerns 5.1 and early DVD; modern streaming specs differ. Method. He describes the film-standard calibration as pink noise at -20 dBFS RMS in one channel at 83 dB SPL (C-weighted, slow) and says authorities recommended lowering monitor gain by 6 dB for home mixes because home rooms do not accommodate high SPLs and wide dynamic range. Example. He says that with a wide-range theatre mix on DVD, home listeners complain it is too loud, or lose dialogue when they turn the volume down. Failure modes. At reduced monitor gain soft passages become too soft, so either compress (about 6 dB of upward compression) or use decoder metadata or an optional home compressor. Limitations. Katz states 83 dB SPL, while later standards use 85 dBC per channel, as discussed in C3-B743-06; Netflix's near-field figures (79 to ...
• articleWhat sound levels do Dolby's Atmos cinema specifications require speakers to reach, and what do they say about reference playback level?
  Problem. Reference level (85 dBC) is a calibration setpoint, but a cinema also needs headroom; agents confuse the two. Conditions. Dolby Atmos cinemas; Issue 5, October 2025. Method. The specification requires each screen speaker system and each surround array to reach 105 dB maximum continuous SPL at the reference listening position (RLP), two-thirds of the way to the rear wall, and 99 dB for individual speakers, with 3 dB amplifier headroom recommended. Example. Inference: against an 85 dBC per-channel reference, 105 dB would be about 20 dB of headroom; the specification does not state the 85 dBC reference in the sections I read. Failure modes. Undersized speakers or amplifiers force operators to lower the fader or clip. Limitations. The document says it is impossible to state with certainty the speaker requirements for achieving standard cinema levels in all cases because of B-chain processing and screen loss. SMPTE ST 202 and RP 200 (calibration standards) and the Dolby Model 73...
• articleIs there a legal or regulatory loudness limit for cinema audiences or adverts in the UK, and what do cinema operators use instead?
  Problem. People ask whether a statutory cinema loudness (or cinema-ad loudness) limit exists. Conditions. UK; the trade body page is dated 2023-01-12. Method. The UK Cinema Association states that there is no statutory standard for audience noise exposure and that cinemas use the 2005 Noise at Work Regulations criteria as a guide; the Regulations set lower action values of 80 dB(A) daily/weekly exposure and 135 dB(C) peak, upper action values of 85 dB(A) and 137 dB(C), and exposure limit values of 87 dB(A) and 140 dB(C). Failure modes. These are employee exposure values, not audience limits, and are A/C-weighted, not Leq(m); do not equate 85 dB(A) with 85 dB Leq(m). Limitations. No UK or other region's cinema-advertising loudness rule was found in fetchable primary sources (the UK Cinema Association page covers sound levels generally; ad-specific rules from the advertising bodies or Ofcom were unreachable or 403). Version/date. UKCA page 2023-01-12; SI 2005/1643 as made.
• articleIs there a published way to describe how loud a whole feature film is, and does it set a limit?
  Problem. A single Leq(m) over a feature hides sustained loud reels, but peak limits would constrain artistic intent. Conditions. DCP-derived measurements of selected features, 2007-2015; the paper is explicitly descriptive. Method. Take a rolling 10-minute Leq(m) with the M-weighting, sampling about every 750 ms, integrate only values above a threshold, and divide by the film duration; the threshold started at 85 dB and the author concluded 80 dB gave the best hierarchy. Example. The paper's table ranks U2 3D highest (528) down to Birdman (77) and Grand Torino (0), with Mad Max: Fury Road at 422. Failure modes. Slumdog Millionaire scored low at an 85 dB threshold because its energy sat between 80 and 85 dB Leq(m). Limitations. The paper says the metric has no absolute value and only ranks films; it does not regulate features; the window length and threshold are subjective choices. Version/date. ISDCF-Doc11 (Leq(m10) 2.0), 2016-03-15.
• articleWhat is the TASA trailer loudness limit in Leq(m), when did it start and how much quieter did trailers get?
  Problem. Trailers in the 1990s were much louder than features and drove fader reductions. Conditions. US feature trailers rated through the MPAA; the account is Ioan Allen's, 2015. Method. The paper says TASA adopted Leq(m) as the measure, tied approval to the MPAA ratings process (administration from June 1999, no rating without TASA certification), started at a maximum of 87 dB Leq(m), lowered it in two steps to 85 dB Leq(m) in 2001, and kept it for both film and digital cinema. Example. It reports the loudest 1996-97 trailers measured 92 and 93 dB Leq(m), so the loudest trailers were about 8 dB quieter after TASA; over 6000 US trailers had passed by 2015. Failure modes. The limit applies to trailers, not to features or to advertising outside the TASA route. Limitations. The 85 figure is in dB Leq(m), not dBC or dB SPL; the current TASA rule text was not fetched (the TASA site did not resolve for me), so this is a 2015 snapshot. Version/date. ISDCF-Doc11, 2016-03-15.
• articleWhat is Leq(m), what does the M in it stand for, and why is it used for trailers and cinema ads rather than A- or C-weighting?
  Problem. Cinema trailers and ads need a loudness number that tracks perceived loudness, and peak or A/C-weighted levels did not. Conditions. Short-duration promotional material with strong upper-mid content; the account is from Dolby's Ioan Allen. Method. Leq(m) is an Leq (long-term average level) with an M-weighting curve, proposed after Dolby research in 1996; the paper says M-weighting, favouring the 2 to 3 kHz region, gave a better loudness 'hierarchy' (ranking of louder/quieter) on trailers and cinema-specific commercials than C-weighting, A-weighting or none. Example. The paper's 'm' stands for 'movies'. Failure modes. A plain long-term Leq on a whole feature misleads (a loud final reel averages low), which is why a rolling window was later proposed in C3-B784-04. Limitations. The M-weighting curve values are in a figure, not the text; the Leq(m) meter/standard documents (ISO 21727, Dolby meter manuals) were not fetched, so the exact filter is not stated here. Version/date. IS...
• articleWhat does 'fader 7' mean in a cinema, and why do many theatres play features below reference (and what does that do to dialogue)?
  Problem. Mixers and agents assume a theatrical mix is heard as mixed, but cinemas often run quieter. Conditions. Applies to cinemas aligned to the Dolby/SMPTE fader-7 reference; the source is a 2015-16 paper by Dolby's Ioan Allen, so it is a manufacturer account, not an audit. Method. The paper states fader 7 as the calibrated reference setting that reproduces the dubbing-stage level; it says theatre faders were lowered to 6, 5 or 4.5, up to about 10 dB below reference, after trailer loudness complaints, and that where projection had no fader automation the reduced trailer level carried over to the feature, hurting dialogue intelligibility. Example. The paper names 1990s trailers (Twister, Star Wars re-release, Men in Black) as the complaint cases. Failure modes. Mixing with heavy sustained loudness invites operators to turn the fader down, which cuts dialogue as well as effects. Limitations. The paper does not state the 85 dBC figure in the text I read, and gives no survey of how m...
• articleWhat does the BBC Writersroom radio scene-style format say about character count, presence, proximity cues and music in sound-only drama?
  Problem: a writer or an AI script generator must convey who is present, how close they are and when the scene changes when the audience has no picture. Conditions: scripts for a sound-only drama; the source is a one-document format sample for the BBC's house radio scene style, not a full craft course. Method (as the document states): (1) a character is established only by speaking or being referred to by name, so every character must either talk or be named; (2) the document warns that too many characters in one scene makes the listener lose track; (3) technical direction should be sparing and describe only the immediate sound pattern, with no novelistic text; (4) music is generally only written in when it is a source cue or does a job such as a transition between scenes; (5) parentheticals such as (OFF), (LOW), (CLOSE), (D) for device-distorted voice and (V.O.) tell the actor how to be close to or far from the microphone. Example from the document: CLOSE means close proximity to th...
• articleHow does CTC forced alignment turn a transcript into word timestamps (frames, blanks, merge, score), and is the torchaudio API still available?
  Problem: you have an accurate transcript and need word start/end times for subtitles, dubbing timing or text-based editing. Conditions: CTC-trained acoustic model (tutorial uses the multilingual MMS alignment bundle). Method: compute the emission (frame-wise token probabilities); call forced_align with the tokenised transcript and blank index 0; the result is in the emission's frame coordinates with blanks and repeats; merge_tokens collapses repeats into token spans with scores; group tokens into words; word score is the average weighted by span length; convert frames to seconds with the ratio of waveform samples to emission frames. A special star token handles missing transcript. Failure modes: Inference: without the star token, untranscribed speech must be absorbed by neighbouring transcript tokens (the tutorial only says the star token handles missing transcript), and low word scores flag mismatched text. Limitations: the page itself warns that torchaudio moved to maintenance in ...
• articleIn an EEG experiment, how did listeners respond to an AI-voiced versus human-voiced newscast, and does it show lower credibility?
  Corrected 2026-10-23 - see C3-B715-02; the original statements below are kept for history and are superseded where the correction says so. Problem. Whether synthetic voice reduces the credibility of narrated information. Conditions. 30 participants (16 female, 14 male, mostly 18 to 40), Chinese-language news items read by an AI voice or a human reporter, with emotional or neutral content; EEG and ERP measures. Measured result. Brain activity responses were greater for the human newscast than for the AI newscast, with differences in beta bands in the left posterior temporal lobe and a content-valence by voice-agent interaction. Source statement vs inference. The authors interpret the larger effect for human voice as possibly reflecting trust in its credibility and fluency; no behavioural credibility rating is the basis of that statement. Limitations. Authors state the sample is small and age and gender were not strictly controlled; one language and one TTS system; EEG differences do ...
• articleDoes a more enthusiastic synthetic narrator improve learning from a narrated lesson?
  Problem. Whether adding emotional cues to a TTS narrator helps viewers learn. Conditions. Between-subjects online experiment, n = 244 learners at a large Asian university, listening to Amazon Alexa narrating a multimedia lesson on a denial-of-service attack in a neutral, low-, medium- or high-enthusiasm voice. Measured result. Enthusiastic voices did not raise immediate retention or transfer scores over the neutral voice; medium and high voices were perceived as more enthusiastic; low and high enthusiasm lowered intrinsic and extraneous cognitive-load ratings; the medium voice raised positive emotion from baseline; persona ratings did not improve. Failure modes. Strong prosody can add load for non-native listeners (the authors' argument). Limitations. One 2022 voice product, one topic, university students, immediate tests only, no human-voice control, no words-per-minute figures extracted. Inference (labelled): richer TTS expressiveness is not shown to improve retention, so do not j...
• articleDo listeners rate cloned or generic AI voices as less trustworthy than human voices, and can they tell them apart?
  Problem. Whether an AI narrator voice is judged as real and trustworthy. Conditions. UK listeners rated unfamiliar voices on short stimuli in rating tasks (Experiments 1a and 1b, with generic AI voices and voice clones in separate contexts) and in a combined Experiment 2 (n = 50; 40 human voices, 40 clones, 40 generic AI voices). Measured result. Generic AI voices were labelled human on 41% of trials and voice clones on 58%; AI voices were not hyperreal (generic voices were rated less real and human than human voices; clones about as real as human voices). Both types were rated more dominant; generic AI voices were rated more trustworthy than human voices in Experiment 1a (57.9 vs 53.8, p = .002) while clones were not significantly different (58.9 vs 57.1, p = .133). Failure modes. Realness ratings of human voices depended on the context of other voices heard. Limitations. First-impression traits on single short stimuli, unfamiliar identities, UK listeners; says nothing about longer...
• articleDoes 1.25x to 1.5x playback change quiz scores, satisfaction or how the speaker is judged, in preregistered online experiments?
  Problem. Whether a slightly faster pace changes viewer experience, speaker impressions and learning. Conditions. Three preregistered online experiments: Study 1 a tiramisu tutorial at 1x, 1.25x or 1.5x (final N = 326 after 53% exclusions); Study 2a an economics video at 1x vs 1.25x with a quiz-warning manipulation (N = 313 after 54.4% exclusions); Study 2b the same in undergraduates (N = 246 after 25.2% exclusions). Measured result. Faster playback reliably raised perceived audiovisual distortions and modestly reduced satisfaction; in Study 1 viewers were more satisfied at 1x (M = 4.17) than at 1.25x (M = 3.82, p = 0.011); the speaker was rated more positively at 1x (M = 4.44) than at 1.25x (4.24) or 1.5x (4.20), marginally significant (p = 0.055), and in Study 2b 1x beat 1.25x on speaker rating (p = 0.022). Objective quiz performance did not differ by speed. Failure modes. Very high exclusion rates; self-report scales; one video per content type. Limitations. Speed is a playback mu...
• articleIn a small randomized within-subject test, how much worse is recall of a story heard at 2x speed than at normal speed?
  Problem. Whether double-speed audio costs memory. Conditions. 40 Lebanese university students (aged 18 to 25, convenience sample), each heard two four-sentence stories of 23 seconds, one at normal speed and one at 2x (after a habituation to 2x words), in counterbalanced order, then played Minesweeper for 10 minutes before cued recall. Measured result. Participants recalled about half of the information overall; normal speed gave a 27% improvement in cued recall over time-compressed speech, statistically significant. Students also reported using compressed speech more than average and mostly judged it harmful. Failure modes. Participants admitted rehearsing during the filler game; mood, time of day and location were uncontrolled. Limitations. Very short material, one narrator, one compression ratio (2x), delayed recall only after 10 minutes, no comprehension test, no ideal-pace data; cannot say what a 10 to 20% faster read would do. The authors suggest faster speeds for revision rath...
• articleHow much does speeding up narration or lecture video hurt learning, according to a 2025 meta-analysis, and how reliable is it?
  Problem. Choosing a narration pace (or a playback speed) for instructional or explainer video. Conditions. Evidence is about accelerated instructional video with different teacher appearance, subtitles and subject matter, mostly educational samples in China and the US. Method (what was measured). A meta-analysis of 12 experimental or quasi-experimental papers, converting speed to words per minute and pooling Hedges' g: acceleration raised cognitive load (g = 0.59, nine studies), lowered maintenance test scores (g = -0.41), lowered transfer ('migration') scores (g = -0.50) and lowered learning satisfaction (g = -0.24). Moderators: teacher on screen, subtitles, country and subject changed some effects; the review reports speeds from 1.1x to 2.5x across the pooled studies, and notes the satisfaction drop arose in samples whose original speech was already 183 to 251 wpm. Failure modes. Heterogeneity is large for cognitive load (I2 77.8%) and satisfaction (I2 96.4%); the authors find mis...
• articleCORRECTION to C3-B054-G19-07: which parts of the sound-planning candidate are in its cited pages (Haunting about 100 sounds and two months; Matrix concept from the finished script; Welles memo) and which are not (Thom 2000 questions; late hiring and composer coordination)?
  Problem. C3-B054-G19-07 lists four planning artefacts and says sound people are 'typically brought on late', spotting is squeezed and composer and sound crew 'rarely coordinate', as 'Weis reports'. Audit result (cited pages re-fetched and searched whole-document). (a) Supported: The Haunting article says Rydstrom had two months to prepare, 'an unusually long time', against sound being viewed as the last and hastiest element; De Bont operated a digital audio player and used about 100 of Rydstrom's sounds on set to inspire and surprise the actors. (a) The Matrix interview: Davis says the Wachowskis had a finished script which they read and 'we developed an overall concept for its sound design'. (a) Welles's sound memo to Murch separates realistic 'background music' from sparingly used 'underscoring'; the page dates it 1958 (the original said 1957-58). (c)/(d) Not supported by the cited pages: the Thom (2000) script-reading questions (no Thom article is among the cited URLs; 'Thom' app...
• articleCORRECTION to C3-B054-G19-06: do the cited FilmSound pages really contain a Weis claim that absence stands out in dense tracks, and Murch 'coronas' for characters? (No: those two parts are not in the cited pages; the silence and signature-sound parts are.)
  Problem. C3-B054-G19-06 says on sound's narrative roles, silence and motifs that 'Weis adds that in dense tracks absence stands out' and that motifs include 'Murch's character coronas'. Audit result (fidelity audit, 23 Oct 2026, four cited filmsound.org pages re-fetched and searched whole-document): supported parts. (a) The three roles direct narrative, subliminal and grammatical are on the Roles of Sound page, which is an edited excerpt of Tomlinson Holman's 1997 introduction to Sound for Film and TV, not Weis. (a) Murch calls silence useful but 'not used that often' and cites the end of The Godfather Part II, where he dropped the track to the atmos and then shut it down. (a) Murch cites Welles's held silence of 20 or 30 seconds in the War of the Worlds radio broadcast (the page says 'War of the Worlds'; the original candidate's '1938 radio drama' date is not stated on the cited page). (a) Signature sounds per character: Murch says the main characters' cars in American Graffiti eac...
• articleWhich parts of a match-action cut (motion before, motion after, sound) create edit blindness (Smith & Martin-Portugues Santacreu 2017)?
  Problem: editors say a cut should start on the action; which side of the cut matters? Conditions: Hollywood clips containing match-action cuts. Method (as reported): motion was removed before the cut, after it, or both, with or without the original soundtrack, while viewers tried to detect cuts; lab eye-tracking with 32 recruited, 27 analysed participants (age 19-35); an online replication ran 80 clips across 8 conditions with 8 participants per clip-condition. Results: removing post-cut motion made cuts quicker to detect and less likely to be missed; removing pre-cut motion had no effect; viewers reoriented less to new content across intact match-action cuts; audio played a large part and edit blindness mostly disappeared without audio. Mapping to AI shot design (inference): when joining two generated clips, the second clip's first frames should already contain the motion that continues the first, and a continuous audio bed is part of the cut, not decoration; a silent preview can m...
• articleWhat is worldizing, how is it actually done (equipment layout, tracks recorded, how they are mixed), and why use it instead of reverb?
  Corrected 2026-10-23 - see C3-B683-02; the original statements below are kept for history and are superseded where the correction says so. Three voices describe the same method: play a finished recording through a speaker in a real space and re-record it with a microphone. Murch's American Graffiti account (quoted on a filmsound.org glossary page from a published interview) gives the concrete layout: a speaker rotated while a microphone about fifty feet away is moved, two passes, then three tracks (dry plus two worldized) blended in the mix to move between present and distant. Welles's 1957-58 sound notes, as quoted in a 1998 article, specify a cheap horn in an alley and say an echo chamber plus filter is not enough. The stated purpose is perceptual: diffuse music leaves dialogue as the focus.
• articleHow do I keep timecode, frame rate and the tempo map consistent across a DAW, a notation program and picture revisions?
  Corrected 2026-10-23 - see C3-B683-03; the original statements below are kept for history and are superseded where the correction says so. Vendor and practitioner texts show that the picture reference carries a start timecode and frame rate that must be set consistently, and that the tempo map can be moved between tools as a MIDI file: Dorico 2.2 (Nov 2018) added tempo-track import and export (timecode offset, tempo changes, time signatures, markers) explicitly for updated edits; Sibelius 2024.3 added pasting a system selection as MIDI per a 2024 article; the orchestrator S10 returns a new tempo map whenever tempos are retimed and trims measure numbers while marking the sketch "Trimmed". Sources document the interchange but not a procedure for re-conforming music to a changed edit.
• articlesox_ng (a4a14b4f) is a hard fork of sox-14.4.2 hosted on Codeberg and by default installs as sox_ng/libsox_ng, not replacing 'sox' unless built with --enable-replace, so scripts calling 'sox' may still invoke the...
  Corrected 2026-10-23 - see C3-B691-01; the original statements below are kept for history and are superseded where the correction says so. sox_ng (a4a14b4f) is a hard fork of sox-14.4.2 hosted on Codeberg and by default installs as sox_ng/libsox_ng, not replacing 'sox' unless built with --enable-replace, so scripts calling 'sox' may still invoke the original.
• articleHow do sound designers layer one hero sound from several components, and what goes wrong when they do?
  Corrected 2026-10-23 - see C3-B683-01; the original statements below are kept for history and are superseded where the correction says so. Five practitioner voices (Serafine/Hall/Watters on The Hunt for Red October 1990, Kirchberger, Rydstrom, Davis, Stambler) describe the same pattern: gather unlike sources that share a quality of the target, process them (pitch, speed, reverse, delay, harmonizer), stack them, add a low or distant sweetener, and sequence the result to picture. The recorded failures are specific: purely synthetic sonar pings were rejected by the director; real animal recordings did not match animatronic mouth movement; literal sources can be unconvincing (Murch).
• articleIs the Welles Touch of Evil sound-notes article cited in C3-B054-G19-03 from 1998, as that candidate says?
  Correction (b, overstated date) to C3-B054-G19-03. Problem: the candidate says Welles's 1957-58 sound notes are 'quoted in a 1998 article'. The fetched page states the article was originally published in Videography Magazine, January 1999; 1998 is the date of the re-edit it discusses, not the article. The method claim itself is supported: the notes say an echo chamber with filter is not enough and call for a cheap horn in an alley. Method: cite the article as Videography, January 1999. Limitations: the printed notes' own date (1957-58) is stated by the candidate and was not independently confirmed from the page text beyond the 1957 film reference.
• articleDoes C3-B054-G19-02 name sources and a Murch point that its four cited pages do not contain?
  Correction (d, unsupported) to C3-B054-G19-02. Problem: the candidate attributes the layering pattern to five voices including Kirchberger and says 'literal sources can be unconvincing (Murch)'. Re-fetching the four cited filmsound.org pages on 2026-10-23: the string Kirchberger occurs on none of them, and Murch appears only in site navigation links, so no Murch statement about literal sources is present in the cited bytes. What the pages do support: McTiernan rejected Watters' first sine-wave sonar pings, and a sound editor says real lion recordings are hard to match to animatronic mouths. Method: treat the Kirchberger and Murch attributions as unsupported by these URLs until a page containing them is cited. Limitations: the claims may exist on other filmsound.org pages; this check covers only the four listed URLs.
• articleWhat does IASA TC-03 (4th edition) specify for analogue-to-digital transfer, replay speed/equalisation, preservation master format and the split between preservation and restoration/access copies?
  Corrected 2026-10-23 - see C3-B645-07; the original statements below are kept for history and are superseded where the correction says so. For preservation transfers IASA-TC 03 (2017) requires replay settings (speed, playback EQ, track format) chosen objectively from knowledge of the original format and documented, forbids subjective 'improvements' such as de-noising in the unmodified transfer, allows playback EQ to be deferred to access copies, and recommends at least 48 kHz/24-bit LPCM, with 96 kHz/24-bit widely used by memory institutions.
• articleWhich ISDCF-naming, hearing-impaired (HI), visually-impaired (VI) and language fields does ClairMeta compare with the CPL, and which does it only parse?
  Corrected 2026-10-23 - see C3-B645-06; the original statements below are kept for history and are superseded where the correction says so. ClairMeta (b3b66a02) parses the DCP title against its ISDCF 9.6 rule set, including HI and VI flags and the audio language, but no check module uses those parsed values: its 12 title-claim checks compare frame rate, dimension, aspect ratio, subtitle and caption presence, channel count, immersive sound, resolution, standard, Dolby Vision, EclairColor and package type with CPL or probe data. The only language comparison is subtitle XML versus CPL. All dcnc checks report at INFO by default, so a default run does not treat title/CPL mismatches as errors.
• articleWhat do AC-3 dialnorm, Android's MPEG-4/MPEG-D DRC decoder keys and Apple's iTunNORM tag document about stream-carried loudness metadata, and what do players do with it?
  Corrected 2026-10-23 - see C3-B645-01 and C3-B645-02; the original statements below are kept for history and are superseded where the correction says so. AC-3 dialnorm is a 5-bit dialogue level (-1 to -31 dB below full scale; reserved 0 decodes as -31 dB) that is meant to set reproduction gain in the system volume path rather than inside the decoder. Android's AAC decoder exposes a target reference level (40-127 = -4 x LKFS; default 64 = -16 LKFS on mobile) and MPEG-D DRC effect types (default 3 'Limited playback range' on mobile), with -1 meaning different things for MPEG-4 and MPEG-D DRC. mpv defaults AC-3 DRC to off despite the standard's default. Apple's iTunNORM is known only from a reverse-engineered 2013 wiki.
• articleI want to transcribe speech or translate it into English with Whisper: which model and task should I use, how many languages, and what do the repository and the Hugging Face card say about the licence?
  Corrected 2026-10-23 - see C3-B645-04; the original statements below are kept for history and are superseded where the correction says so. At whisper 86098128 the README states that Whisper's code and model weights are released under MIT (LICENSE is MIT text), while the Hugging Face openai/whisper-large-v3 card (06f233fe) carries 'license: apache-2.0' as metadata; both model cards say strong ASR in about ten languages and caution against transcribing people without consent.
• articleAs a newcomer choosing an open TTS model for the voice stage of a dubbing pipeline, which common ones have permissive weights, which are non-commercial or custom-licensed, and what do they say about languages?
  Corrected 2026-10-23 - see C3-B644-03; the original statements below are kept for history and are superseded where the correction says so. Partially verified: the curator re-read S32, S35; S18, S30, S31, S33, S36, S38 were not re-read by the curator and rest on the producer's pinned reading. Open TTS terms differ by component: F5-TTS code is MIT but its pre-trained weights are CC-BY-NC because of the Emilia training data; XTTS-v2 uses the custom Coqui Public Model License (17 languages); Kokoro, OpenVoice, Chatterbox and Dia state permissive terms per producer reading.
• articleDo the cited IASA page and current iasa-web.org URLs support the TC-03 replay-parameter, deferred-EQ and BWF 48/96 kHz claims?
  Problem. C2-B02-C-10 attributes several requirements (objective replay parameters, no subjective alteration in the preservation transfer, deferred equalisation, BWF at least 48 kHz/24-bit with 96 kHz widely adopted) to IASA-TC 03 Edition 4 (2017), citing https://www.iasa-web.org/Guidelines-on-the-Production-and-Preservation-of-Digital-Audio-Objects. Conditions. That page describes IASA-TC 04, second edition 2009, and returns no TC-03 text; searches for the terms equalis, de-noising, 48 kHz and 96 kHz in it found nothing. Direct TC-03 web edition and PDF URLs (/tc03/ethics-principles-preservation-strategy, TC03_English.pdf, both found via search) returned 404 on 2026-10-23. Method. Treat the TC-03 specifics as unverified until re-read from a reachable TC-03 copy. Failure modes. Misattributing TC-04 (2009) content to TC-03 (2017), or citing numeric formats with no reachable source. Limitations. The claims may well be accurate to the original TC-03 text the earlier worker read; this au...
• articleWhich tasks does the Whisper repository document beyond transcription and translation to English?
  Problem. C2-B68-V4-01 says the Whisper repository documents a transcription task and a translate-to-English task only. Conditions. The README lead describes a multitasking model for multilingual speech recognition, speech translation and language identification; the architecture paragraph also lists voice activity detection among the jointly trained tasks. The command-line --task option exposes transcribe and translate. Method. Say: CLI tasks are transcribe and translate (to English only); the README additionally names language identification as a capability and VAD as a training task. Failure modes. An agent would skip language detection when the model can provide it. Limitations. This audit did not run Whisper. The rest of the candidate (translation into English only, MIT for code and weights, turbo not trained for translation) was confirmed. Version/date. openai/whisper at the pinned commit, fetched 2026-10-23.
• articleIs there a Dutch rule of at most 4.5 LU between voice loudness and programme loudness?
  Problem. C2-B27-C4-09 reports a Dutch rule of at most 4.5 LU between voice loudness and programme loudness. Conditions. EBU Tech Review (June 2022) section 4.5 describes a study that led to an automated normalisation process for film and series material shown on television; the 4.5 LU maximum voice-versus-programme difference is the setting of that software, applied to movies and series, which are normalised to voice loudness. Method. Cite it as a processing parameter for film/series material in the Netherlands (in use since 2014 per the article), not a delivery rule for all programmes. Failure modes. Applying 4.5 LU to factual or live content as if it were a requirement. Limitations. Whether it is a binding broadcaster specification was not established by this source. Version/date. EBU Technical Review, June 2022, fetched 2026-10-23.
• articleIs AC-3 dialnorm applied by the decoder or by the system volume path according to ATSC A/52:2012?
  Problem. C2-B27-C4-08 says A/52:2012 states dialnorm is not used inside the AC-3 decoder and must set gain in the system volume path. Conditions. Section 5.4.2.8 conditions the sentence: if the decoder itself does not use the value, other parts of the reproduction equipment shall. Section 7.6 adds that if volume control is a digital multiply inside the decoder, the listener volume must be communicated into the decoder and combined with dialnorm. Method. Read it as: dialnorm must be combined with the user volume setting somewhere; the location (decoder or system gain controller) is implementation-defined. Failure modes. Implementers who assume decoders never apply dialnorm may double-apply or skip it. Limitations. Only A/52:2012 text was checked; later A/52 revisions not read. Version/date. ATSC A/52:2012 PDF, fetched 2026-10-23.
• articleDoes Android's KEY_AAC_DRC_TARGET_REFERENCE_LEVEL (40-127, default 64) apply to AC-3/dialnorm content?
  Problem. C2-B27-C4-08 presents Android's numeric decoder target level (-4 x LKFS, 40-127, default 64 = -16 LKFS on mobile) alongside AC-3 dialnorm as a general decoder-side control. Conditions. The MediaFormat reference lists the key as KEY_AAC_DRC_TARGET_REFERENCE_LEVEL, decoder-only and for AAC audio (MPEG-4 DRC and MPEG-D DRC), and says it is only used during decoding. Method. Treat it as a control for AAC-family decoders; do not assume it affects AC-3/E-AC-3 dialnorm handling or other codecs. Failure modes. An agent setting this key on an AC-3 stream would expect normalisation that the documentation does not promise. Limitations. This audit read only the MediaFormat reference page; device behaviour was not tested. Version/date. Android MediaFormat reference page fetched 2026-10-23.
• articleWhere does the Chatterbox TTS licence appear: Hugging Face card prose, card metadata, or the GitHub repository?
  Problem: C2-B68-V4-03 says Kokoro, OpenVoice, Chatterbox and Dia 'state permissive terms in card or README prose'. Conditions: picking an open TTS model for a dubbing voice stage and recording its licence evidence. Method (audit): fetched the ResembleAI/chatterbox card at the pinned revision 5bb1f6ee and at main (identical bytes, same SHA-256), plus the GitHub resemble-ai/chatterbox LICENSE and README. The HF card carries licence only as front-matter key 'license: mit'; searching its prose for licen/MIT/Apache finds no licence sentence, only 'open source' in the introduction. The MIT text is in the GitHub repository LICENSE. Kokoro (prose 'Apache-licensed'), Dia (prose 'Apache License 2.0') and OpenVoice V2 (prose 'MIT License') do have prose statements. Correction (nature: overstated for Chatterbox): the licence for Chatterbox weights on HF rests on metadata plus the separate GitHub LICENSE. Also the card now presents Multilingual V3 with '23+' languages and a 'Single Language Pack...
• articleHow does Rhubarb time mouth shapes relative to audio, what does it do for non-English speech, and who owns the output?
  Problem. Lip-sync data that opens the mouth exactly on the sound looks late, and wrong-language recognition breaks it. Conditions. Rhubarb 2D lip-sync pipeline. Method. Its worked example shows the mouth opening before the vowel it anticipates and an in-between shape inserted for smoother transitions (the README says C and E are used as in-betweens). Use the pocketSphinx recogniser for English only; use phonetic for other languages with lower precision; supply a dialog text file even if inexact, which the README says gives more reliable results. Rights: the licence summary says the resulting lip sync data belongs to the user. Failure modes. Wrong recogniser for the language. Limitations. Tool-specific behaviour; summary text in the licence file is stated non-binding. Date. 2026-10-23.
• articleWhat is the practical difference between Rhubarb's mouth-shape set and Preston Blair's, and what lip-sync checks follow for AI-generated 2D or 3D talking characters?
  Problem. Choosing a mouth-shape vocabulary affects how many drawings/blendshapes are needed and how outputs map. Conditions. 2D or stylised lip-sync from recorded dialogue. Method. Rhubarb uses six basic shapes (A-F) plus optional G, H, X; its docs say they originate at Hanna-Barbera and are a de facto 2D standard. For Moho/OpenToonz .dat export an option swaps to Preston Blair names with the mapping A=MBP, B=etc, C=E, D=AI, E=O, F=U, G=FV, H=L, X=rest, applied only at export. Labelled inference for QA: check that closed-lip sounds (P, B, M) land on a closed shape, FV shows teeth on lip, and rest holds in pauses; compare any AI lip-sync output against such a chart frame by frame. Failure modes. Name mapping done after animation; OpenToonz needs Blair names. Limitations. Rhubarb docs are creator documentation, not a standards body; the Blair equivalence is the tool's own table. Date. 2026-10-23.
• articleDoes C3-B095-G35-01 mis-describe the one differing ElevenLabs clause 4(h), given that the EEA version also keeps the HIPAA bar?
  Problem. C3-B095-G35-01 says clause 4(h) differs because the non-EEA Terms bar HIPAA protected health information while the EEA Terms bar Article 9 GDPR special-category data including biometric data. Conditions. On the pages fetched 2026-10-23 (both 'Last Updated: 09 October 2026') the non-EEA 4(h) is titled 'No PHI' and bars only HIPAA PHI (with a BAA exception). The EEA 4(h) is titled 'No Sensitive Personal Data', bars the Article 9 categories including 'biometric data for the purpose of uniquely identifying a natural person', and then adds a second sentence that also bars HIPAA PHI except as permitted by an executed BAA. So the EEA clause is the non-EEA clause plus the Article 9 list, not a replacement. Method. When comparing the two Terms, describe EEA 4(h) as a superset. Failure modes. Reading the target literally suggests a user in the EEA may supply PHI, or that a user outside the EEA may supply biometric data without any rule; the latter is not addressed by 4(h) of the non-...
• articleDoes C3-B116-G42-03 leave out that commercial audio detectors reached 0.89 accuracy in Deepfake-Eval-2024, and why the authors discount it?
  Problem. C3-B116-G42-03 summarises commercial detectors as better than open-source ones but 'none reached 90 percent accuracy', and then gives only video (0.78) and image (0.82 to 0.86) numbers. Conditions. Deepfake-Eval-2024 (arXiv 2503.02857v5) Table 5 (best commercial model per modality) reads video 0.78 accuracy / 0.79 AUC, audio 0.89 / 0.93, image 0.82 / 0.90 (image precision 0.99, recall 0.71). Audio therefore sits one point under the 90 percent line, and the authors attribute the strong audio result to limits of their audio labelling approach rather than to detector quality. Method. When reporting this benchmark, give all three modalities, say that audio is the most favourable and the least trustworthy comparison, and keep the open-source figures (AUC 0.43 to 0.63 off the shelf; real-media false-positive rates 0.31 to 0.39 video, 0.22 to 0.99 image, up to 1.00 for audio P3 in Table S8) alongside. Failure modes. Omitting audio makes the commercial range look uniformly weak (ab...

More