At 6 kbps, Meta’s EnCodec neural codec scored 83.1 on the MUSHRA listening scale for clean speech, while Opus at the same 6 kbps scored 30.1 and the hidden original scored 95.5 (Meta AI, EnCodec paper, TMLR 2023). That gap is the whole story of neural audio coding in one row: the same bits, a score 53 points higher (83.1 minus 30.1). Google’s SoundStream at 3 kbps matched quality that EVS needed at least 9.6 kbps and Opus at least 12 kbps to reach (Google, SoundStream paper 2021), and research codecs now push usable speech below 1 kbps, down to 0.16 kbps (Della Libera et al., FocalCodec, NeurIPS 2025). Classic codecs are answering with neural add-ons of their own: Opus 1.5 made speech “perfectly usable down to 6 kb/s” with a 400 MFLOPS enhancer (Xiph.Org, Opus 1.5 release 2024), which matters for anyone tuning call quality alongside jitter buffers and POLQA scores. We aggregated data from Google, Meta AI, Descript, Kyutai, Xiph.Org, the IETF, 3GPP, the ITU, and peer-reviewed benchmarks such as Codec-SUPERB, DASB, WavTokenizer, TAAE and FocalCodec.
TL;DR
- EnCodec at 6 kbps scored 83.1 MUSHRA on clean speech vs 30.1 for Opus at 6 kbps (Meta AI, EnCodec paper 2023).
- SoundStream at 3 kbps matched quality that needed at least 9.6 kbps for EVS and 12 kbps for Opus (Google, SoundStream paper 2021).
- Lyra V2 at 3.2 kbps was rated on par with Opus at 10 kbps in MUSHRA tests (Google Open Source Blog, Lyra V2 2022).
- The Descript Audio Codec compresses 44.1 kHz audio about 90x into 8 kbps tokens (Descript, DAC paper, NeurIPS 2023).
- FocalCodec at 0.16 kbps reached a UTMOS of 4.22 against 4.09 for the uncompressed reference (Della Libera et al., NeurIPS 2025).
- The TAAE codec delivered a PESQ of 2.96 at 400 bps, close to Mimi’s 3.01 at 1,100 bps (Parker et al., ICLR 2025).
- VALL-E learned zero-shot TTS from 60K hours of English speech coded as neural codec tokens, needing only a 3-second voice prompt (Microsoft, VALL-E paper 2023).
- Opus 1.5’s NoLACE enhancer at 9 kb/s beat non-enhanced Opus at 12 kb/s, at about 0.75% CPU (Xiph.Org, Opus 1.5 release 2024).
- Kyutai’s Mimi codes speech at 12.5 Hz and 1.1 kbps, enabling a dialogue model with 200 ms practical latency (Kyutai, Moshi paper 2024).
- EnCodec at 1.5 kbps more than tripled ASR word error rate, from 2.96% to 9.21% (Wu et al., Codec-SUPERB 2024).
1. Bits Needed to Match Classic Codecs in Listening Tests
The headline result of the first neural codec generation was not “better sound”, it was “the same sound for a fraction of the bits”. SoundStream at 3 kbps reached a quality level that EVS needed at least 9.6 kbps and Opus at least 12 kbps to match, which works out to 3.2x and 4x the bits (9.6 / 3 = 3.2; 12 / 3 = 4) (Google, SoundStream paper). The advantage narrows as the budget grows: at medium and high bitrates the classic codecs needed 2.2x to 2.6x and then only 1.3x to 1.6x more bits.
MUSHRA, the listening test used in most of these papers, is defined in ITU-R BS.1534-3: listeners rate several coded versions of a clip side by side with a hidden copy of the original, so a codec’s score is best read against what the hidden original scored in the same test. The pattern is consistent across vendors: the lower the bitrate, the larger the neural lead, partly because Opus at 6 kbps falls back to narrowband audio.
| Metric | Value | Source |
|---|---|---|
| Opus and EVS bitrate needed to match SoundStream at 3 kbps | Opus at least 12 kbps, EVS at least 9.6 kbps | Google, SoundStream paper (arXiv 2021) |
| Extra bits EVS and Opus need to match SoundStream at medium bitrate (6 kbps) | 2.2x to 2.6x more | Google, SoundStream paper (arXiv 2021) |
| Opus bitrate rated on par with Lyra V2 at 3.2, 6 and 9.2 kbps (MUSHRA) | 10, 13 and 14 kbps | Google Open Source Blog, Lyra V2 (2022) |
| MUSHRA, clean speech: EnCodec at 6 kbps vs Opus at 6 kbps vs hidden original | 83.1 vs 30.1 vs 95.5 | Meta AI, EnCodec paper (TMLR 2023) |
| MUSHRA, music set 1: EnCodec at 3 kbps vs Opus at 12 kbps | 89.6 vs 77.8 | Meta AI, EnCodec paper (TMLR 2023) |
| POLQA MOS at 6 kbps wideband, MLow vs Opus (vendor’s own claim) | 3.9 vs 1.89 | Meta Engineering, MLow blog (2024) |
Most recent available data: SoundStream (2021), Lyra V2 (2022) and EnCodec (2022 preprint, 2023 journal). These are still the reference comparisons cited by newer papers, but they predate Opus 1.5’s neural enhancements (see section 7), so the classic-codec baselines are older than today’s libopus. The MLow figures come from Meta’s engineering blog, which notes that Opus runs in narrowband mode (0 to 4 kHz) at 6 kbps; it is a vendor’s own claim from a production calling codec, included as a reference point rather than an independent measurement.
2. Compression Ratios and Where Quality Stops Improving
Compression ratio is where neural codecs look most dramatic. The Descript Audio Codec (DAC) compresses 44.1 kHz audio about 90x into tokens at 8 kbps with one universal model for speech, music and environmental sound (Descript, DAC paper, NeurIPS 2023). In 48 kHz stereo music, EnCodec at 6 kbps (a 256x compression) scored 82.9 MUSHRA, the same perceived quality band as MP3 at 64 kbps (24x compression, 82.7).
The more useful finding for engineers is saturation. EnCodec’s stereo model scored 88.0 at 12 kbps and 87.5 at 24 kbps, so doubling the bitrate bought nothing measurable. A second lever is entropy coding: a small language model over the tokens trims 25% to 40% of the bandwidth with no quality change, at the price of speed.
| Metric | Value | Source |
|---|---|---|
| Compression of 44.1 kHz audio by DAC at 8 kbps | About 90x | Descript, DAC paper (NeurIPS 2023) |
| ViSQOL at 44 kHz: DAC at 8 kbps vs EnCodec at 24 kbps vs Opus at 24 kbps | 4.18 vs 3.16 vs 4.15 | Descript, DAC paper (NeurIPS 2023) |
| ViSQOL at 44 kHz: DAC at 1.78 kbps vs Opus at 8 kbps | 3.76 vs 2.06 | Descript, DAC paper (NeurIPS 2023) |
| MUSHRA, 48 kHz stereo music: EnCodec at 6 kbps (256x) vs 12 kbps vs 24 kbps vs MP3 at 64 kbps (24x) | 82.9 vs 88.0 vs 87.5 vs 82.7 | Meta AI, EnCodec paper (TMLR 2023) |
| Bandwidth saved by a language model used as entropy coder on EnCodec tokens | About 25% to 40% | Meta AI, EnCodec paper (TMLR 2023) |
| Potential rate saving from entropy coding SoundStream tokens | 7% to 20% | Google, SoundStream paper (arXiv 2021) |
ViSQOL is an objective model that predicts perceived quality from the signal; DAC’s own MUSHRA scores are published only as plots, so the table uses its printed ViSQOL figures. One caveat on the Opus comparison: in the DAC paper Opus at 8 kbps was coded with only 4 kHz of audio bandwidth, while DAC kept the full 22.05 kHz, so part of the gap is bandwidth, not efficiency.
3. The Sub-1 kbps Frontier for Speech
The 2025 generation of research codecs moved the fight from “beat Opus” to “how few bits can carry intelligible, natural speech”. FocalCodec at 0.16 kbps scored a UTMOS of 4.22 on clean LibriSpeech, above the 4.09 of the uncompressed reference (Della Libera et al., FocalCodec, NeurIPS 2025). UTMOS is a neural model that predicts a listener’s mean opinion score, so a codec can “beat” the original on it simply by producing very clean-sounding output.
Intelligibility tells the other half. The same table reports dWER, the extra word error rate an ASR system makes on the resynthesized speech: FocalCodec’s 0.16 kbps setting reached 7.94, while its 0.65 kbps setting held to 2.18. Natural sound is now cheap; keeping every word and the speaker’s identity still costs bits.
| Metric | Value | Source |
|---|---|---|
| UTMOS: FocalCodec at 0.16 kbps vs uncompressed reference (LibriSpeech clean) | 4.22 vs 4.09 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| dWER: FocalCodec at 0.65 kbps vs at 0.16 kbps | 2.18 vs 7.94 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| UTMOS and dWER: BigCodec at 1.04 kbps | 4.11 and 2.55 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| UTMOS and dWER: EnCodec at 1.5 kbps | 1.58 and 8.08 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| PESQ: TAAE at 400 bps vs Mimi at 1,100 bps | 2.96 vs 3.01 | Parker et al., TAAE paper (ICLR 2025) |
| PESQ and SI-SDR: TAAE at 700 bps | 3.09 and 4.73 dB | Parker et al., TAAE paper (ICLR 2025) |
| UTMOS: WavTokenizer at 0.9 kbps (75 tokens per second) vs DAC at 9 kbps vs ground truth | 4.0486 vs 3.9097 vs 4.0562 | Ji et al., WavTokenizer paper (ICLR 2025) |
The TAAE comparison is worth reading twice: at 400 bps it reached roughly Mimi’s PESQ at 1,100 bps, a codec that is itself only a few years old (Parker et al., TAAE paper). Note that every figure in this section is measured on clean LibriSpeech or LibriTTS read speech, so treat it as a best case.
4. Latency and Compute Cost
Bitrate is only one budget; a codec for live calls or streaming also has a delay budget and a CPU budget. Lyra V2 encodes and decodes a 20 ms frame in 0.57 ms on a Pixel 6 Pro, 35 times faster than real time, with 20 ms of latency versus 100 ms for Lyra V1 (Google Open Source Blog, Lyra V2). Streaming neural codecs that work in 20 ms frames fit inside the delay range the Opus standard already allows for classic coding (see the table).
Model size is where the generations diverge. The codecs aimed at phones stay small and fast, while the sub-1 kbps research codecs trade compute for bits: TAAE has about 950M parameters, and SemantiCodec runs slower than real time.
| Metric | Value | Source |
|---|---|---|
| Latency: Lyra V2 vs Lyra V1 | 20 ms vs 100 ms | Google Open Source Blog, Lyra V2 (2022) |
| Real-time factor on one MacBook Pro 2019 CPU thread at 6 kbps, encoder and decoder (above 1 = faster than real time): EnCodec 24 kHz vs Lyra v2 | 9.8 and 10.4 vs 27.4 and 67.2 | Meta AI, EnCodec paper (TMLR 2023) |
| Opus algorithmic delay range and frame sizes | 5 to 65.2 ms; frames of 2.5 to 60 ms | IETF, RFC 6716 (2012) |
| Model parameters: DAC (encoder plus decoder) | 76M (22M encoder, 54M decoder) | Descript, DAC paper (NeurIPS 2023) |
| Model parameters: TAAE vs Mimi vs EnCodec | 950M vs 80M vs 14M | Parker et al., TAAE paper (ICLR 2025) |
| Real-time factor (above 1 = faster than real time): FocalCodec at 0.16 kbps vs BigCodec vs SemantiCodec | 208 vs 22 vs 0.62 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
Real-time factor is reported differently across papers: EnCodec and FocalCodec use “times faster than real time” (higher is better), so values from different papers should not be compared row to row. The Opus range comes from the IETF’s RFC 6716, the 2012 standard that still defines the Opus bitstream; most recent available data for that baseline is the 2012 RFC, extended by the 2024 and 2025 libopus releases covered in section 7.
5. Codec Tokens in Speech and Audio Language Models
Neural codecs now have a second career: they are the tokenizers that let language models speak. Kyutai’s Mimi codes 24 kHz speech into tokens at 12.5 Hz and 1.1 kbps, using 8 codebooks of 2,048 entries, and the Moshi dialogue model built on it reports a theoretical latency of 160 ms and 200 ms in practice (Kyutai, Moshi paper). An autoregressive model generally runs one step per token frame, so frame rate translates directly into speed, a theme covered from the agent side in our time-to-first-audio-byte statistics.
That is why frame rate, not just bitrate, became the design target. NVIDIA’s Low Frame-rate Speech Codec runs at 21.5 frames per second and 1.89 kbps and, per its authors, makes LLM-based text-to-speech inference around three times faster. Voice products built on voice cloning software and TTS inherit these trade-offs directly.
| Metric | Value | Source |
|---|---|---|
| Mimi token rate, bitrate and audio sample rate | 12.5 Hz, 1.1 kbps, 24 kHz | Kyutai, Moshi paper (arXiv 2024) |
| MUSHRA: Mimi at 1.1 kbps vs SpeechTokenizer at 4.0 kbps vs SemantiCodec at 1.3 kbps | 81.0 vs 74.3 vs 64.8 | Kyutai, Moshi paper (arXiv 2024) |
| Moshi dialogue latency, theoretical vs in practice | 160 ms vs 200 ms | Kyutai, Moshi paper (arXiv 2024) |
| Frame rate and bitrate of NVIDIA’s Low Frame-rate Speech Codec | 21.5 frames per second, 1.89 kbps | NVIDIA, Low Frame-rate Speech Codec paper (arXiv 2024) |
| Speed-up of LLM text-to-speech inference with that low-frame-rate codec | Around 3x faster | NVIDIA, Low Frame-rate Speech Codec paper (arXiv 2024) |
| Tokens per second of 24 kHz audio with WavTokenizer’s single quantizer | 40 or 75 | Ji et al., WavTokenizer paper (ICLR 2025) |
| Training data and voice prompt for VALL-E, a TTS model trained on neural codec codes | 60K hours of English speech; 3-second prompt | Microsoft, VALL-E paper (arXiv 2023) |
| MusicGen tokenizer: EnCodec at 32 kHz, codebooks, token rate, training music | 4 codebooks of 2,048, 50 Hz, 20K hours | Copet et al., MusicGen paper (arXiv 2023) |
Most recent available data for VALL-E and MusicGen: 2023. A caveat for the MUSHRA row: Mimi’s 81.0 is the variant trained with adversarial losses only; the variant trained without them scored 58.8, so the tokenizer’s training recipe moves perceived quality as much as its bitrate does.
6. What Codecs Preserve for Downstream Tasks
Listening scores do not tell you whether a transcription engine, a speaker check or an emotion classifier still works on the decoded audio. In Codec-SUPERB, EnCodec at 1.5 kbps raised ASR word error rate from 2.96% to 9.21% and speaker-verification equal error rate from 0.86% to 13.88%, while DAC at 24 kbps kept word error rate at exactly the original 2.96% (Wu et al., Codec-SUPERB). Those error-rate shifts matter for anyone tracking speech recognition hallucination and error rates in compressed pipelines.
Speaker identity and emotion are the first casualties of low bitrates: speech content survives longer than the paralinguistic layer. The DASB benchmark reaches a sharper conclusion for codec tokens used as model inputs: “Across all domains and tasks, continuous representations outperform discrete tokens”, and medium bitrates worked best (Mousavi et al., DASB).
| Metric | Value | Source |
|---|---|---|
| Codec models and model families evaluated in Codec-SUPERB | 19 variants from 6 families | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| ASR word error rate: original audio vs EnCodec at 1.5 kbps | 2.96% vs 9.21% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| Speaker verification equal error rate: original audio vs EnCodec at 1.5 kbps | 0.86% vs 13.88% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| Emotion recognition accuracy: original audio vs EnCodec at 1.5 kbps | 69.84% vs 58.84% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| ASR word error rate: DAC at 6 kbps vs DAC at 24 kbps | 3.26% vs 2.96% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| Correlation of PESQ with the overall speech score | 0.97 | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| DASB bitrate formula for tokenizers | log2(vocabulary size) x codebooks x token rate | Mousavi et al., DASB benchmark (arXiv 2024) |
Context note: equal error rate is the point at which a speaker-verification system wrongly accepts and wrongly rejects at the same rate, so a jump from 0.86% to 13.88% means the decoded voice is much harder to match to its owner. Codec-SUPERB’s PESQ correlation also suggests a practical shortcut: for speech, the long-established PESQ metric still tracks the benchmark’s overall verdict closely.
7. Classic Codecs Go Neural: Opus, EVS and the Standards Track
The classic codecs have not stood still. Opus 1.5 (March 4, 2024) showed that its NoLACE enhancer makes 9 kb/s speech better than non-enhanced 12 kb/s, at a run-time cost of about 0.75% CPU (Xiph.Org, Opus 1.5 release). Its Deep Redundancy (DRED) feature, now on the IETF standards track, packs up to one second of backup audio into each packet, so that “every 20-ms packet is effectively transmitted 50 times”, and the release notes say it “still works well even with 90% loss” (here “packet loss” means data packets dropped on the network, not lost audio files).
The upshot: the gap measured in section 1 against older Opus builds is narrowing from the classic side. Opus 1.6 (December 15, 2025) went further, shrinking the DRED model and adding neural bandwidth extension so that 9 kb/s speech reaches the quality of Opus 1.4 fullband at 18 kb/s.
| Metric | Value | Source |
|---|---|---|
| Opus bitrate range defined by the standard | 6 kbit/s to 510 kbit/s | IETF, RFC 6716 (2012) |
| EVS bitrate range and maximum audio bandwidth | 5.9 to 128 kbps, up to 20 kHz | 3GPP, EVS codec news release (2014) |
| Lowest Opus 1.5 bitrate at which speech is “perfectly usable” with NoLACE | 6 kb/s | Xiph.Org, Opus 1.5 release (2024) |
| Complexity of the NoLACE and LACE enhancers and the FARGAN neural vocoder in Opus 1.5 | NoLACE 400 MFLOPS (about 0.75% CPU); LACE 100 MFLOPS (about 0.15% CPU); FARGAN 600 MFLOPS (1/5 of LPCNet) | Xiph.Org, Opus 1.5 release (2024) |
| DRED model size: Opus 1.6 vs Opus 1.5 | 600 kB vs 1,800 kB | Xiph.Org, Opus 1.6 release (2025) |
| Opus 1.6 bandwidth extension with NoLACE at 9 kb/s matches | Opus 1.4 fullband at 18 kb/s | Xiph.Org, Opus 1.6 release (2025) |
| DRED redundancy bitrate vs regular Opus bitrate, and in-band LBRR for comparison | About 1/50 (DRED) vs about 2/3 (LBRR) | IETF, draft-ietf-mlcodec-opus-dred-07 (August 4, 2026) |
The DRED specification is still an Internet-Draft (IETF mlcodec working group), revised on August 4, 2026; the Opus 1.5 notes warned that “the version included in Opus 1.5 will not be compatible with the final version”. EVS figures are from the 3GPP Release 12 announcement; most recent available data: 3GPP, 2014.
Summary: Neural Audio Codec Bitrate vs Quality by the Numbers
| Metric | Value | Source |
|---|---|---|
| MUSHRA, clean speech: EnCodec at 6 kbps vs Opus at 6 kbps | 83.1 vs 30.1 | Meta AI, EnCodec paper (TMLR 2023) |
| Opus and EVS bitrate needed to match SoundStream at 3 kbps | Opus at least 12 kbps, EVS at least 9.6 kbps | Google, SoundStream paper (arXiv 2021) |
| Opus bitrate range defined by the standard | 6 to 510 kbit/s | IETF, RFC 6716 (2012) |
| DAC compression of 44.1 kHz audio at 8 kbps | About 90x | Descript, DAC paper (NeurIPS 2023) |
| ViSQOL: DAC at 8 kbps vs EnCodec at 24 kbps | 4.18 vs 3.16 | Descript, DAC paper (NeurIPS 2023) |
| dWER: FocalCodec at 0.65 kbps vs at 0.16 kbps | 2.18 vs 7.94 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| UTMOS: FocalCodec at 0.16 kbps vs reference | 4.22 vs 4.09 | Della Libera et al., FocalCodec paper (NeurIPS 2025) |
| PESQ: TAAE at 400 bps vs Mimi at 1,100 bps | 2.96 vs 3.01 | Parker et al., TAAE paper (ICLR 2025) |
| UTMOS: WavTokenizer at 0.9 kbps vs DAC at 9 kbps | 4.0486 vs 3.9097 | Ji et al., WavTokenizer paper (ICLR 2025) |
| Emotion recognition accuracy: original vs EnCodec at 1.5 kbps | 69.84% vs 58.84% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| Model parameters: TAAE vs EnCodec | 950M vs 14M | Parker et al., TAAE paper (ICLR 2025) |
| Mimi token rate and bitrate | 12.5 Hz, 1.1 kbps | Kyutai, Moshi paper (arXiv 2024) |
| Moshi dialogue latency in practice | 200 ms | Kyutai, Moshi paper (arXiv 2024) |
| LLM text-to-speech speed-up from a 21.5 frames per second codec | Around 3x | NVIDIA, Low Frame-rate Speech Codec paper (arXiv 2024) |
| ASR word error rate: original vs EnCodec at 1.5 kbps | 2.96% vs 9.21% | Wu et al., Codec-SUPERB paper (arXiv 2024) |
| MUSHRA: Mimi at 1.1 kbps vs SpeechTokenizer at 4.0 kbps | 81.0 vs 74.3 | Kyutai, Moshi paper (arXiv 2024) |
| NoLACE-enhanced Opus at 9 kb/s compared with non-enhanced Opus | Better than 12 kb/s | Xiph.Org, Opus 1.5 release (2024) |
| DRED model size, Opus 1.6 vs Opus 1.5 | 600 kB vs 1,800 kB | Xiph.Org, Opus 1.6 release (2025) |
Methodology and Sources
- Google, SoundStream: An End-to-End Neural Audio Codec (arXiv 2107.03312, 2021; IEEE/ACM TASLP): MUSHRA bitrate ratios vs Opus and EVS, frame size, entropy-coding savings.
- Google Open Source Blog, Lyra V2 (September 2022): Opus equivalence, latency, Pixel 6 Pro timing.
- Meta AI, High Fidelity Neural Audio Compression (EnCodec), TMLR 2023 (arXiv 2210.13438): MUSHRA tables, compression ratios, real-time factors, entropy coding.
- Meta Engineering, MLow: Meta’s low bitrate audio codec (June 13, 2024): vendor’s own POLQA comparison with Opus.
- Copet et al., Simple and Controllable Music Generation (MusicGen), arXiv 2023: EnCodec tokenizer configuration and training hours.
- Descript, High-Fidelity Audio Compression with Improved RVQGAN (DAC), NeurIPS 2023: compression factor, ViSQOL table, parameter count.
- Kyutai, Moshi: a speech-text foundation model for real-time dialogue (arXiv 2410.00037, 2024): Mimi codec specification, MUSHRA vs semantic tokenizers, dialogue latency.
- NVIDIA, Low Frame-rate Speech Codec (arXiv 2409.12117, 2024): frame rate, bitrate, TTS inference speed-up.
- Microsoft, VALL-E: Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (arXiv 2301.02111, 2023): training hours and prompt length.
- Ji et al., WavTokenizer (ICLR 2025, arXiv 2408.16532): single-quantizer token rates and UTMOS.
- Parker et al., Scaling Transformers for Low-Bitrate High-Quality Speech Coding (TAAE), ICLR 2025: 400 and 700 bps results, parameter counts.
- Della Libera et al., FocalCodec (NeurIPS 2025, arXiv 2502.04465): sub-1 kbps benchmark table with UTMOS, dWER and real-time factors.
- Wu et al., Codec-SUPERB: An In-Depth Analysis of Sound Codec Models (arXiv 2402.13071, 2024): downstream ASR, speaker, emotion results.
- Mousavi et al., DASB: Discrete Audio and Speech Benchmark (arXiv 2406.14294, revised April 21, 2026): discrete vs continuous representation findings.
- Xiph.Org, Opus 1.5 release notes (March 4, 2024) and Opus 1.6 release notes (December 15, 2025): LACE, NoLACE, DRED, FARGAN, bandwidth extension.
- IETF, RFC 6716: Definition of the Opus Audio Codec (September 2012) and IETF, draft-ietf-mlcodec-opus-dred: Opus bitrate and delay ranges, DRED overhead.
- 3GPP, Enhanced Voice Services codec for LTE (Release 12, November 7, 2014): EVS bitrate range and bandwidth.
- ITU-R BS.1534-3 (MUSHRA), October 2015 and ITU-T P.808 crowdsourced speech quality, June 2021: the listening-test methods behind the MUSHRA and crowdsourced MOS figures.
- Data watch: (1) MUSHRA, MOS, ViSQOL, UTMOS and PESQ are not interchangeable, and a score is only meaningful against the reference and baselines in the same test; UTMOS in particular can rate a codec above the original (FocalCodec 4.22 vs 4.09). (2) Meta’s MLow blog prints “POLQA MOS 1.89 vs 3.9” without labelling which score belongs to which codec; the article context implies Opus 1.89 and MLow 3.9, and the Opus version used is not stated. (3) Codec-SUPERB’s text says DAC 24 kbps emotion accuracy drops “by 0.38%”, but its table values (69.84% to 69.56%) give 0.28 points, and the paper’s dataset count varies between 20 and 21; we report only the table values. (4) WavTokenizer’s 40-token setting is printed as 0.5 kbps in its table and 0.48 kbps in its text. (5) EnCodec’s arXiv page still says “Preprint” while the TMLR 2023 listing confirms journal publication; Google lists SoundStream in IEEE/ACM TASLP with a 2021 date while other indexes give 2022. (6) Parameter counts for the same codec vary by source: Mimi is listed at 80M in the TAAE paper, 82M in the FocalCodec paper and 96.2M on Kyutai’s Hugging Face model card, likely depending on which components are counted. Real-time factor definitions also differ between papers, so cross-paper speed rows are not directly comparable. (7) Most head-to-head comparisons with Opus predate the Opus 1.5 and 1.6 neural features, and no market-size figures are included because no Tier 1 market data for neural codecs was found.
- Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the IETF mlcodec working group advances draft-ietf-mlcodec-opus-dred toward RFC publication.