--- license: other license_name: webai-non-commercial-license-ver.-1.0 license_link: https://huggingface.co/webAI-Official/Echo-Air/blob/main/LICENSE.md base_model: gradium/phonon-31m-en base_model_relation: quantized pipeline_tag: text-to-speech library_name: gguf language: - en tags: - gguf - text-to-speech - tts - phonon - ptts - on-device - cpu --- # Echo-Air A 4-bit build of Gradium's [phonon-31m-en](https://huggingface.co/gradium/phonon-31m-en) English text-to-speech model. It is a drop-in replacement for Gradium's own `model.q4k.gguf`: the same 55,075,168 bytes, the same file name, and it loads in Gradium's unmodified `ptts` runtime on CPU. In a pre-registered test on 2,576 synthesized clips, its speech was **as intelligible as the BF16 model's** (word error rate within 0.5 points) and **closer to BF16's audio than Gradium's Q4_K file**. It also scored higher predicted quality than that file. > **Non-commercial use only. Released under the webAI Non-Commercial License ver. 1.0**. ### At a glance Passes all four rules of the pre-registered test. | | BF16 (Gradium) | Gradium Q4_K | **Echo-Air** | |---|---:|---:|---:| | File size | 102,823,920 B | 55,075,168 B | **55,075,168 B** | | Word error rate, all 1,273 sentences | 1.83% | 2.02% | **1.91%** | | Word error rate, 200 hard sentences (numbers, dates, codes) | 5.90% | 6.30% | **5.98%** | | Speaker similarity to BF16's audio of the same sentence | 1 (itself) | 0.9859 | **0.9866** | | Predicted quality (UTMOS) | 4.333 | 4.281 | **4.314** | | Latent error vs BF16, teacher-forced screen (400 items, separate protocol) | 0 (itself) | 0.01608 | **0.01284** | Audio rows: stock ptts 0.6.0 on CPU, temperature 0.3, one seed per clip shared by every file, 1,273 sentences × 2 voices, one run per file. Word error rate from Whisper large-v3. See [Evaluation protocol](#evaluation-protocol). ![Word error rate, distance from BF16's audio and UTMOS for BF16, Gradium's Q4_K and Echo-Air](figures/benchmarks.png) ## Highlights * **As intelligible as BF16.** Word error rate 1.91% vs BF16's 1.83%, a difference of +0.08 points (95% CI −0.02 to +0.18), inside the pre-registered +0.5-point margin. No suite is significantly worse. Gradium's Q4_K is significantly worse than BF16 on the same test: +0.19 points (+0.06 to +0.32). * **Closer to BF16's voice than Gradium's Q4_K.** Speaker similarity to BF16's clip of the same sentence is 0.9866 vs 0.9859, +0.0007 (95% CI +0.0002 to +0.0011), and higher on both suites. * **Higher predicted quality than Gradium's Q4_K.** UTMOS 4.314 vs 4.281, +0.033 (95% CI +0.027 to +0.039). * **Same size, same runtime.** A byte-for-byte drop-in for `model.q4k.gguf` in Gradium's repository; nothing else changes. It is **not identical to BF16**: its audio is still slightly further from BF16's than BF16 resampled with a second seed, and its UTMOS is 0.019 lower than BF16's. See [Limitations](#limitations). ## Model Details | Property | Value | |---|---| | Developed by | webAI Intelligence Lab | | Model ID | `webAI-Official/Echo-Air` (proposed; not yet published) | | Base model | [`gradium/phonon-31m-en`](https://huggingface.co/gradium/phonon-31m-en) at revision `92741c5b`, 31M parameters plus the Mimi audio decoder | | Model file | `model.q4k.gguf`, 55,075,168 bytes (sha256 `ef1b486d7072d3db…`) | | Format | GGUF. The 24 transformer matrices of the flow LM are 4-bit Q4_K, the same tensor types and layout as Gradium's `model.q4k.gguf`; every other tensor is unchanged | | Runs on | Gradium's `ptts` runtime, CPU backend (`device="cpu"`, `quant="q4k"`). Tested with ptts 0.6.0 on Linux x86-64 and on macOS (Apple silicon) | | Input / output | English text and a voice / mono 24 kHz audio | | Voices | Gradium's [phonon-voices](https://huggingface.co/gradium/phonon-voices) catalog and the included `default-voice.safetensors` | | License | webAI Non-Commercial License ver. 1.0 | ## Results Every audio result comes from one test, run once: each file synthesized the same 2,576 clips (1,273 distinct sentences in 2 voices) with stock ptts 0.6.0 on CPU, temperature 0.3, and the same seed per clip for every file. Verdicts pair sentences (each sentence's 2 voices are averaged first) and use a bootstrap 95% interval. "BF16" is Gradium's `model.safetensors`; "Gradium Q8_0" and "Gradium Q4_K" are the two GGUF files in Gradium's repository; "llama.cpp Q4_K" is a second same-size file, the same matrices quantized to Q4_K with llama.cpp's standard quantizer; the two "BF16" control rows rerun BF16 with the same seeds (runtime noise) and with a second seed (how much BF16 differs from itself when resampled). ### Pre-registered test Echo-Air passes all four rules, which were written down before the test ran. | Check | Rule | Result (95% CI) | Pass | |---|---|---|---| | Primary: Echo-Air vs BF16, word error rate | upper end of the 95% interval below +0.005, and no suite significantly worse | 0.0183 → 0.0191, +0.0008 (−0.0002 to +0.0018) | yes | | Primary: Echo-Air vs Gradium Q4_K, similarity to BF16 | lower end of the 95% interval above 0 | 0.9859 → 0.9866, +0.0007 (+0.0002 to +0.0011) | yes | | Guardrail: Echo-Air vs Gradium Q4_K, word error rate | not significantly worse | 0.0202 → 0.0191, −0.0011 (−0.0024 to +0.0002) | yes | | Guardrail: Echo-Air vs Gradium Q4_K, UTMOS | lower end of the 95% interval above −0.05 | 4.2814 → 4.3141, +0.0327 (+0.0267 to +0.0388) | yes | These four results are confirmatory. Everything else on this page (per-suite rows outside the rules, the control rows, and the screen) is supporting detail and should be read as preliminary. ### Word error rate Echo-Air shows no significant change from BF16, while both other 4-bit files are significantly worse than BF16. | File | Size | All (1,273) | Seed-TTS English (1,073) | Hard (200) | Corpus WER | vs BF16, all, points (95% CI) | Verdict | |---|---:|---:|---:|---:|---:|---|---| | BF16 (baseline) | 102,823,920 B | 1.83% | 1.07% | 5.90% | 1.82% | baseline | | | BF16, rerun with the same seeds | same file | 1.83% | 1.07% | 5.90% | 1.81% | −0.00 (−0.01 to +0.00) | no significant change | | BF16, second seed | same file | 1.82% | 1.10% | 5.67% | 1.81% | −0.01 (−0.13 to +0.09) | no significant change | | Gradium Q8_0 | 64,512,352 B | 1.82% | 1.04% | 6.05% | 1.81% | −0.01 (−0.08 to +0.06) | no significant change | | Gradium Q4_K | 55,075,168 B | 2.02% | 1.23% | 6.30% | 1.98% | +0.19 (+0.06 to +0.32) | higher (worse) | | llama.cpp Q4_K | 55,075,168 B | 1.96% | 1.14% | 6.31% | 1.93% | +0.13 (+0.03 to +0.22) | higher (worse) | | **Echo-Air** | 55,075,168 B | **1.91%** | **1.16%** | **5.98%** | **1.88%** | +0.08 (−0.02 to +0.18) | no significant change | The columns give the mean per-sentence word error rate (each clip capped at 1; no clip reached the cap), with the number of sentences in brackets; "Corpus WER" is total errors over total words. Per suite against BF16, Echo-Air is +0.09 points on Seed-TTS English (95% CI −0.00 to +0.18) and +0.08 on the hard sentences (−0.37 to +0.50), neither significant; the Seed-TTS interval just touches zero. Against Gradium's Q4_K, Echo-Air is −0.11 points (−0.24 to +0.02), no significant change. The two BF16 controls show the test's own noise: the same-seed rerun differs only by transcription noise, and a second seed moves the score by −0.01 points. ![Word error rate per suite for every file](figures/wer.png) ### Speaker similarity to BF16's audio Echo-Air is closer to BF16's audio than both other 4-bit files, but not as close as BF16 resampled. | File | All | Seed-TTS English | Hard | vs Gradium Q4_K, all (95% CI) | Verdict | |---|---:|---:|---:|---|---| | BF16, rerun with the same seeds | 1.0000 | 1.0000 | 1.0000 | +0.0141 (+0.0136 to +0.0145) | higher | | Gradium Q8_0 | 0.9956 | 0.9956 | 0.9959 | +0.0097 (+0.0093 to +0.0101) | higher | | BF16, second seed | 0.9872 | 0.9866 | 0.9906 | +0.0013 (+0.0009 to +0.0018) | higher | | Gradium Q4_K (baseline) | 0.9859 | 0.9855 | 0.9880 | baseline | | | llama.cpp Q4_K | 0.9859 | 0.9853 | 0.9887 | −0.0000 (−0.0005 to +0.0004) | no significant change | | **Echo-Air** | **0.9866** | **0.9861** | **0.9890** | +0.0007 (+0.0002 to +0.0011) | higher | Similarity is the cosine between speaker-verification embeddings of a file's clip and BF16's clip of the same sentence, voice and seed. Echo-Air beats Gradium's Q4_K on both suites: Seed-TTS English +0.0006 (+0.0001 to +0.0011) and hard +0.0010 (+0.0001 to +0.0020). Against BF16 resampled with a second seed it is −0.0007 (−0.0011 to −0.0002), significantly lower. ![Similarity to BF16's audio per file, and each file's difference from Gradium's Q4_K](figures/similarity.png) ### Predicted quality (UTMOS) Echo-Air scores higher than Gradium's Q4_K but lower than BF16. | File | All | Seed-TTS English | Hard | vs Gradium Q4_K, all (95% CI) | Verdict | |---|---:|---:|---:|---|---| | BF16 | 4.333 | 4.328 | 4.361 | +0.052 (+0.046 to +0.058) | higher | | BF16, rerun with the same seeds | 4.333 | 4.328 | 4.361 | +0.052 (+0.046 to +0.058) | higher | | BF16, second seed | 4.330 | 4.324 | 4.360 | +0.049 (+0.043 to +0.055) | higher | | Gradium Q8_0 | 4.332 | 4.327 | 4.359 | +0.050 (+0.045 to +0.056) | higher | | Gradium Q4_K (baseline) | 4.281 | 4.278 | 4.300 | baseline | | | llama.cpp Q4_K | 4.283 | 4.277 | 4.311 | +0.001 (−0.005 to +0.007) | no significant change | | **Echo-Air** | **4.314** | **4.309** | **4.343** | +0.033 (+0.027 to +0.039) | higher | Against BF16, Echo-Air is −0.019 (−0.025 to −0.014), significantly lower. ### Against Gradium's Q4_K on the teacher-forced screen (separate protocol) Echo-Air's predictions are 20% closer to BF16's than Gradium's Q4_K's at the identical file size. | File (55,075,168 B each) | Latent error | vs Gradium Q4_K (95% CI) | Verdict | End-of-speech logit error | Stop frame changed | |---|---:|---|---|---:|---:| | Gradium Q4_K (baseline) | 0.01608 | baseline | | 0.224 | 23 / 400 | | llama.cpp Q4_K | 0.01390 | −0.00218 (−0.00249 to −0.00189) | lower | 0.195 | 16 / 400 | | **Echo-Air** | **0.01284** | −0.00324 (−0.00357 to −0.00293) | lower | 0.209 | 8 / 400 | This screen does not synthesize audio. Each file's model predicts every next audio latent from BF16's own previous frames, with the same text, voice and noise, and the error is the mean squared difference from BF16's prediction. It covers 400 items (200 sentences × 2 voices), was run once, and uses transformers on CPU in f32, not ptts. Echo-Air is lower on both sentence sources: WikiText-2 (320 items) 0.01640 → 0.01318, −0.00322 (−0.00356 to −0.00290), and hand-written numbers and dates (80 items) 0.01478 → 0.01148, −0.00331 (−0.00406 to −0.00261). Against llama.cpp's Q4_K it is also lower, −0.00106 (−0.00130 to −0.00082), but its end-of-speech logit error is higher than that file's (0.209 vs 0.195). Against Gradium's Q4_K that error is lower, −0.0151 (−0.0176 to −0.0126), and Echo-Air changes the stop frame least often. ![Latent error against BF16 per sentence source](figures/latent_error.png) ## Files | File | What it is | |---|---| | `model.q4k.gguf` | Echo-Air (55,075,168 bytes, sha256 `ef1b486d7072d3dbd4634b6f846cd818e6fa129a819e928da23e7287b0920fba`); modified from Gradium's `model.q4k.gguf` | | `default-voice.safetensors` | Gradium's default voice, unchanged (256,136 bytes, sha256 `dc537478eaadc1a5257fcda2e3501ab81e14c111a5d95a5838f350db2afd0f1a`) | | `config.json`, `generation_config.json`, `processor_config.json`, `chat_template.jinja`, `tokenizer.json`, `tokenizer.5.0.0.json`, `tokenizer.model`, `tokenizer_config.json` | Gradium's configuration and tokenizer files, unchanged from revision `92741c5b`, so the folder runs on its own | | `LICENSE.md` | webAI Non-Commercial License ver. 1.0 | | `apache-2.0-LICENSE.txt` | Apache License 2.0, the license of Gradium's phonon-31m-en | | `figures/` | The charts in this card | Check the downloads with `shasum -a 256 model.q4k.gguf default-voice.safetensors` against the hashes above. ## Quick start ```bash pip install ptts==0.6.0 hf download webAI-Official/Echo-Air --local-dir Echo-Air && cd Echo-Air ptts -m . -l en -q q4k -d cpu -v Easton "Echo-Air read this sentence on a laptop CPU at 6:20 PM on November 8th, 2026." -o speech.wav ``` In Python: ```python import ptts tts = ptts.TTS(".", lang="en", quant="q4k", device="cpu") tts.save("speech.wav", "Hello from Echo-Air.", voice="Easton") ``` * The `ptts` command was run from inside this folder with ptts 0.6.0; `quant="q4k"` selects `model.q4k.gguf`, and `device="cpu"` is the backend this file was tested on. * The evaluation used temperature 0.3 (the model's default) and a fixed seed per sentence; pass `seed=` (Python) or `-s` (command line) for reproducible audio. * Voices other than the default come from Gradium's voice catalog and are downloaded on first use; every voice in [gradium/phonon-voices](https://huggingface.co/gradium/phonon-voices) works by name or ID. * Streaming (`tts.stream(...)`) and the other runtimes in [Gradium's guide](https://huggingface.co/gradium/phonon-31m-en) work as with Gradium's files; only the CPU backend was tested with this file. ## Limitations * **Not identical to BF16.** Its audio is less similar to BF16's (0.9866) than BF16 resampled with a second seed (0.9872), a significant difference of −0.0007, and its UTMOS is 0.019 lower than BF16's (significant). If the closest possible match to BF16 matters more than size, Gradium's Q8_0 (64.5 MB) is closer to BF16 on every measure here. * **Small effects.** The similarity gain over Gradium's Q4_K is +0.0007 on a 0–1 scale; it is significant over 1,273 sentences but small. The word-error-rate improvement over Gradium's Q4_K (−0.11 points) is not significant. * **One run.** Each file was measured once on one test set, with one seed per clip; only BF16 was also resampled. The screen was also run once. * **Scope.** English only, two voices (Easton and Master Aldric), sentences of up to about 10 seconds of audio. Long-form reading, other voices, voice cloning and listening tests were not evaluated. Measures are automatic (speech recognition, speaker embeddings, predicted quality), not human ratings. * **CPU only.** Results hold for ptts's CPU backend. ptts's other backends (Core ML, CUDA, Metal, Vulkan and the browser build) were not tested with this file. * **Speed was not measured** under controlled conditions. The file has the same size and tensor types as Gradium's Q4_K. One example on an Apple silicon Mac: the quick-start command produced 6.3 s of audio in 1.8 s, model loading included. * **Which file.** Results hold for the included `model.q4k.gguf` (sha256 above), byte-identical to the file that was evaluated. ## Evaluation protocol * **Synthesis:** stock ptts 0.6.0 (Linux wheel), CPU backend, 128 single-thread processes on an AMD EPYC 9355 machine (container limited to 13.6 CPUs); temperature 0.3; one seed per clip, derived from the clip's id and shared by every file; the "second seed" control uses a different seed per clip. Voice catalog `gradium/phonon-voices` at revision `0ee17205`. BF16 ran from `model.safetensors` in f32, Gradium Q8_0 as `q8_0`, and the three Q4_K files as `q4k`. Every file rendered all 2,576 clips on CPU. Clips with unusually long audio (≥ 29.5 s, or > 1.5 s per word) numbered 0–4 per file. * **Sentences:** the 1,088 English target sentences of the Seed-TTS evaluation set (Hugging Face mirror `zhaochenyang20/seed-tts-eval` at revision `8f5e1aa2`; 1,073 distinct) and 200 templated sentences full of numbers, dates, money, codes and units (the "hard" suite), each in the voices Easton and Master Aldric: 2,576 clips, 1,273 distinct sentences. * **Word error rate:** Whisper large-v3 (fp16, greedy, English), Whisper's English text normaliser on both reference and transcript, then digit groups split at "." or ":" and "a m"/"p m" joined (so "3:51 p.m." and "3 51 pm" match); word errors counted with jiwer, each clip capped at 1, then averaged over a sentence's 2 voices. Not comparable with Gradium's published LLM-assisted word error rates. * **Speaker similarity:** cosine between `microsoft/wavlm-base-plus-sv` speaker-verification embeddings of a file's clip and BF16's clip of the same sentence, voice and seed, at 16 kHz. * **Predicted quality:** UTMOS (`utmos22_strong` from `tarepan/SpeechMOS` v1.2.0), at 16 kHz. * **Statistics:** every comparison pairs the same sentences (n = 1,273; 1,073 Seed-TTS and 200 hard); 95% intervals from 2,000 bootstrap resamples of the per-sentence differences. "No significant change" means the interval includes zero. * **Pre-registration:** the question, primary measures, margins (+0.005 word error rate against BF16; UTMOS within 0.05 of Gradium's Q4_K) and pass rules were written down before the test ran. * **Screen:** teacher-forced, 160 WikiText-2 test sentences and 40 hand-written sentences of numbers and dates in the voices Hadley and Zoey (400 items); BF16 reference latents from transformers in f32 on CPU, temperature 0.3, one flow step; latent error = mean squared difference of the predicted normalised latents; stop frame = first frame whose end-of-speech logit exceeds 4.0. ## Responsible Use Do not use Echo Air to impersonate a real person, to deceive listeners about who is speaking, or to clone a voice without the speaker’s consent. Tell listeners when speech is synthetic where that matters to them. ## License and attribution Released under the webAI Non-Commercial License ver. 1.0 (LICENSE.md). This distribution incorporates a modified version of Gradium’s gradium/phonon-31m-en. Gradium’s weights, tokenizer and voices are licensed under Apache 2.0 (text kept in apache-2.0-LICENSE.txt) and its runtime is licensed under MIT OR Apache-2.0. Please follow Gradium’s model card for its responsible use guidance. No Gradium sponsorship, affiliation, or endorsement is implied.