FFmpeg Video and Audio Encoding¶

CLI media encoder. Covers codec selection, NVENC hardware acceleration, audio encoding, tempo change with formant preservation, and concat fast paths. FFmpeg 7.x, NVENC Turing+.

Key Facts¶

• -vsync deprecated in FFmpeg 6+; use -fps_mode cfr or -fps_mode vfr instead
• libfaac removed; use native aac or package management-installed libfdk_aac
• -rc constqp -qp X still works but -rc vbr -cq X -b:v 0 is the recommended NVENC pattern
• -preset slow for NVENC is gone; use numeric -preset p5/p6/p7 (p7 = slowest/best)
• -tag:v hvc1 required for HEVC in MP4 when targeting Apple QuickTime/Safari; default hev1 fails
• NVENC Turing+ (RTX 2000+) required for spatial-aq/temporal-aq; Ada (RTX 4000+) for av1_nvenc
• Rubber Band R3 engine requires standalone CLI from breakfastquay.com; ffmpeg's built-in af_rubberband uses the older R2 engine
• yuv420p mandatory for hardware decoder compatibility (phones, TVs, embedded); yuv420p10le archival only
• -movflags +faststart moves moov atom to file start — required for HTTP streaming/seeking before full download

Codec Selection¶

CRF/CQ equivalence (low-motion content: slides, screen recording, talking head):

Low-motion content (lectures, screencasts, slides) tolerates CRF 2-3 higher than general video — static regions benefit more from AQ than extra bitrate.

Scaler choice:

Command Recipes¶

Stream copy (container repack, no quality loss)¶

# MKV → MP4 repack, no re-encode
ffmpeg -i input.mkv -c copy -movflags +faststart output.mp4

libx264 archival master¶

ffmpeg -i input.mp4 \
  -c:v libx264 -preset slow -crf 20 \
  -profile:v high -level 4.1 -pix_fmt yuv420p \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

h264_nvenc bulk batch (Turing+)¶

ffmpeg -i input.mp4 \
  -c:v h264_nvenc -preset p6 -tune hq \
  -rc vbr -cq 24 -b:v 0 -maxrate 6M -bufsize 12M \
  -bf 2 -spatial-aq 1 -temporal-aq 1 -rc-lookahead 32 \
  -profile:v high -pix_fmt yuv420p \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

-bf 2: B-frames for better compression on low-motion content (NVENC may default to 0).
-rc-lookahead 32: 32-frame buffer for better rate control; higher values give diminishing returns.
-multipass fullres: adds ~+10-15% quality at 2x encode time — worth it for archival, skip for bulk.

HEVC archival (libx265, 10-bit for slide/text content)¶

ffmpeg -i input.mp4 \
  -c:v libx265 -preset slow -crf 22 \
  -pix_fmt yuv420p10le -tag:v hvc1 \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

10-bit reduces posterization and banding on gradient slides; hvc1 tag required for Apple compatibility.

hevc_nvenc bulk¶

ffmpeg -i input.mp4 \
  -c:v hevc_nvenc -preset p6 -tune hq \
  -rc vbr -cq 26 -b:v 0 \
  -spatial-aq 1 -temporal-aq 1 -rc-lookahead 32 \
  -pix_fmt yuv420p \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

SVT-AV1 CPU (recommended over libaom-av1)¶

ffmpeg -i input.mp4 \
  -c:v libsvtav1 -preset 6 -crf 28 \
  -svtav1-params tune=0:lookahead=120 \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

tune=0: VOD quality mode. lookahead=120: large lookahead for quality; reduce to 32 for speed.

av1_nvenc (Ada, RTX 4000+)¶

ffmpeg -i input.mp4 \
  -c:v av1_nvenc -preset p6 -tune hq \
  -rc vbr -cq 30 -b:v 0 \
  -spatial-aq 1 -temporal-aq 1 \
  -c:a aac -b:a 128k -ac 1 -ar 48000 \
  -movflags +faststart \
  output.mp4

Resolution downscale (4K → 1080p)¶

# Photographic / talking head — lanczos
ffmpeg -i input.mp4 -vf "scale=1920:-2:flags=lanczos" -c:v h264_nvenc ... output.mp4

# Screen recording / text slides — bicubic (less ringing)
ffmpeg -i input.mp4 -vf "scale=1920:-2:flags=bicubic" -c:v h264_nvenc ... output.mp4

-2: computes height automatically while keeping aspect ratio; ensures even number required by H.264.

Audio: AAC for voice¶

# Native AAC (universal builds) — adequate at 128k
-c:a aac -b:a 128k -ac 1 -ar 48000

# libfdk_aac (Fraunhofer, better quality especially <128k; requires non-free build)
-c:a libfdk_aac -b:a 96k -ac 1 -ar 48000 -afterburner 1

# Verify libfdk_aac availability
ffmpeg -encoders | grep fdk

AAC bitrate guide for mono voice: 64k = acceptable streaming, 96k = distribution minimum, 128k = archival sweet spot, 160k+ = diminishing returns.
HE-AAC (-profile:a aac_he) better below 64k but has weaker player support; use LC-AAC for compatibility.

Speech tempo change with formant preservation (Rubber Band R3)¶

Three-pass pipeline — extract audio, R3 stretch, re-mux with sped-up video:

# Pass 1: extract audio as lossless WAV
ffmpeg -y -i src.mp4 -vn -c:a pcm_s16le -ar 44100 -ac 1 audio.wav

# Pass 2: R3 time-stretch with formant preservation (-F mandatory for 3x speech)
rubberband-r3 -T 3 -F -q audio.wav audio_3x.wav
# -T 3   : 3x tempo (duration = original / 3)
# -F     : formant preserve — prevents gender-shift on extreme ratios
# -q     : quiet mode for batch

# Pass 3a: speed video only (no audio, use NVENC + AQ)
ffmpeg -y -i src.mp4 -an \
  -filter:v "setpts=PTS/3.0" \
  -c:v h264_nvenc -preset p6 -tune hq -cq 24 \
  -bf 2 -spatial-aq 1 -temporal-aq 1 \
  -movflags +faststart video_3x.mp4

# Pass 3b: mux speed-up video + R3 audio
ffmpeg -y -i video_3x.mp4 -i audio_3x.wav \
  -map 0:v:0 -map 1:a:0 -c:v copy \
  -c:a aac -b:a 128k -ac 1 \
  -shortest \
  -movflags +faststart final_3x.mp4

-shortest aligns stream durations — required when video and audio lengths differ by milliseconds after stretch.

Concat: fast path (no re-encode)¶

All clips must share identical codec, resolution, frame rate, pixel format, and audio parameters:

# Build list file
printf "file '%s'\n" part1.mp4 part2.mp4 part3.mp4 > list.txt

# Demuxer concat — ~37x faster than re-encode
ffmpeg -f concat -safe 0 -i list.txt -c copy output.mp4

# With regenerated timestamps (for speed-up outputs or AVI/MKV rips with corrupt PTS)
ffmpeg -fflags +genpts -f concat -safe 0 -i list.txt -c copy output.mp4

Concat: filter path (re-encode, mixed params)¶

ffmpeg -i part1.mp4 -i part2.mp4 \
  -filter_complex "[0:v][0:a][1:v][1:a]concat=n=2:v=1:a=1[v][a]" \
  -map "[v]" -map "[a]" \
  -c:v h264_nvenc -preset p6 -tune hq -cq 24 \
  -c:a aac -b:a 128k -ac 1 \
  -movflags +faststart output.mp4

Use when clips differ in any parameter; slower but handles any input combination.

NVENC Tuning Reference¶

Gotchas¶

• Issue: af_rubberband=tempo=3.0 (ffmpeg built-in) produces metallic/robotic artifacts at 3x speech. Fix: Use Rubber Band R3 standalone CLI (rubberband-r3 -T 3 -F). Built-in uses R2 engine; even with formant=preserved R2 loses to R3 on speech at high ratios.

• Issue: -movflags +faststart omitted — video buffers entire file before play begins; seeking fails on HTTP streams. Fix: Always append -movflags +faststart for any distribution MP4. No downside for local files.

• Issue: NVENC quality disappoints without -tune hq; default tune is equivalent to unoptimized mode. Fix: Always specify -tune hq for quality target; omitting it leaves significant quality headroom unused with no speed benefit.

• Issue: hevc_nvenc output refuses to play in QuickTime/Safari (hev1 tag not recognized). Fix: Add -tag:v hvc1 to force correct tag; applies to any HEVC in MP4 targeting Apple.

• Issue: ffmpeg -f concat -c copy produces audio/video sync drift when source files have non-zero start PTS (common in speed-up outputs or MKV rips). Fix: Add -fflags +genpts to regenerate timestamps: ffmpeg -fflags +genpts -f concat -safe 0 -i list.txt -c copy out.mp4.

• Issue: CQ too low (aggressive) on NVENC produces visible blocking/ringing on text and slide gradients. Fix: For 1080p lecture/slide content use CQ 24-26 for H.264, CQ 26-28 for HEVC. CQ 18 or below = near-lossless but filesize grows exponentially; unnecessary for screen content.

• Issue: libsvtav1 CRF scale is not the same as x264/x265; direct CRF transplant produces vastly different results. Fix: SVT-AV1 CRF 28-32 ≈ x265 CRF 22 ≈ x264 CRF 16-18 for equivalent perceptual quality on typical content.

• Issue: atempo=3.0 in an ffmpeg audio filter chain produces intelligible but low-quality speech at 3x; phase-vocoder smearing is audible. Fix: Do not use atempo above 2x for speech; use Rubber Band R3 standalone pipeline instead.

See Also¶

• bash scripting - batch processing loops for multi-file encode jobs
• io redirection and pipes - piping ffmpeg output, logging encode progress
• monitoring and performance - GPU utilization monitoring during NVENC encode (nvidia-smi dmon)
• disks and filesystems - storage layout for large video archives
• FFmpeg Codecs Documentation
• FFmpeg Scaler Documentation
• FFmpeg Wiki: Encode/AV1
• FFmpeg Wiki: Encode/AAC
• Rubber Band Audio Library
• NVIDIA Video Codec SDK Presets
• SVT-AV1 Preset Analysis — OTTVerse
• CRF Guide — slhck.info