VoxCPM

Exact model and requirements

Artifact:

OpenBMB / VoxCPM2 native model with AudioVAE V2 auxiliary checkpoint · checkpoint openbmb/VoxCPM2

Publisher revision:

32279effe8c19989596f05d353d1447f51d9e915

Artifact digest:

sha256:f7f964cfa9da23653baec6e6f7750719977ad944ed9f95fe52fe3a620506891d

Download size:

4.27 GiB

Memory requirement (unknown):

Unknown — No full process peak measured for exact native BF16 CPU, FP32 AudioVAE, static caches and output buffers.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor or pipeline sharding. Whole native speech model, text-token preparation, acoustic patch generation and AudioVAE decoding run on CPU; no GPU or disk offload.

KV cache:

Native BF16 static base/residual LM caches, each capacity8192; AudioVAE/state and full process memory unmeasured.

Workload:

8,192 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native LlamaTokenizerFast plus mask_multichar_chinese_tokens; add audio-start control101 · 32279effe8c19989596f05d353d1447f51d9e915; runtime cce58a5b59303c9bd63d12f23afe6a49a4a80c59

Sampling / reasoning:

Native diffusion guidance cfg_value 2.0 and 10 inference steps, min_len 2, max_len 256 acoustic patches, seed 1234, retry_badcase False with one attempt. Native effective cap is min(text_token_count*6+10,256). No text temperature/top-p sampler applies. No chat template or reasoning output. contextTokens 8,192 labels native shared text/acoustic autoregressive positions. Reserve actual textcount+one audio-start+effective acoustic cap within 8,192; native lm_config 32,768 position setting is distinct from allocated cache. Generated acoustic feature count strictly below the effective cap establishes native early stop only for zero-shot context_len 0, not intelligibility/quality.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card explicitly releases VoxCPM2 under Apache-2.0; the native runtime includes the actual Apache-2.0 license text. Retain relevant license/copyright/NOTICE and modification notices; patent and trademark conditions apply. The AudioVAE is part of the same published model repository. This fixed original-text recipe supplies no reference voice or external audio and grants no rights to impersonate anyone. No legal clearance or publisher endorsement is claimed.

• audiovae.pth · 0.35 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• config.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• tokenizer.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• tokenizer_config.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• special_tokens_map.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915

• Main checkpoint 4,580,080,592 bytes plus AudioVAE 376,951,122 bytes = 4,957,031,714 published weight bytes, excluding all runtime/cache/activation/OS overhead.
• 32 GiB RAM, 40 GiB disk and four threads are author planning allocations, not measured minima or guaranteed fit. BF16 CPU support does not imply fast execution.

Install the runtime

1. 1. Create the exact native CPU research environment

   Install exact CPU wheels and pinned direct dependencies, then use the pinned native source checkout via an explicit PYTHONPATH without installing its broad UI/ASR package bundle. The archived runtime import closure supports this fixed zero-shot path. No UI, denoiser, text normalizer, VAD, prompt-audio loading or external model is enabled. Record transitive/system resolution and stop on pip check or imports/model-load errors; this stack is not executed by the author.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl libsndfile1
   mkdir -p "$HOME/voxcpm2-tts/models"
   cd "$HOME/voxcpm2-tts"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91' 'https://download-r2.pytorch.org/whl/cpu/torchaudio-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=dfb1ae1d7da1e869a6a6a315cc2b2652c43e3aabb5184da4d363d1b4bb2c86a4'
   .venv/bin/python -m pip install 'numpy==2.2.3' 'transformers==4.57.6' 'einops==0.8.1' 'pydantic==2.10.6' 'safetensors==0.5.3' 'huggingface-hub==0.36.0' 'librosa==0.11.0' 'soundfile==0.13.1' 'tqdm==4.67.1'
   git clone https://github.com/OpenBMB/VoxCPM.git native-voxcpm
   git -C native-voxcpm checkout --detach cce58a5b59303c9bd63d12f23afe6a49a4a80c59
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned checkout/hashed CPU wheels and direct dependencies resolve; pip check passes for the selected installed packages and resolution is saved. The native source is imported only at reader start, through explicit PYTHONPATH. No actual model load is preclaimed.

Download the model

1. 1. Download and hash all native model and AudioVAE assets

   Reader downloads only; the author archived metadata/source without weight bodies. Use exact native revision for both weights and tokenizer/config files. The manifest pins file bytes rather than mutable API counters. Stop on any failure; do not accept an alternate codec or revision.

   CommandCopy command

   set -e
   cd "$HOME/voxcpm2-tts"
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/model.safetensors' -o 'models/model.safetensors'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/audiovae.pth' -o 'models/audiovae.pth'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/config.json' -o 'models/config.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/tokenizer.json' -o 'models/tokenizer.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/tokenizer_config.json' -o 'models/tokenizer_config.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/special_tokens_map.json' -o 'models/special_tokens_map.json'
   cat > asset-manifest.json <<'PYMANIFEST'
   {
     "model.safetensors": {
       "bytes": 4580080592,
       "sha256": "f7f964cfa9da23653baec6e6f7750719977ad944ed9f95fe52fe3a620506891d"
     },
     "audiovae.pth": {
       "bytes": 376951122,
       "sha256": "94b5d51e107e0507d4acc976cfdadb64edd6fd06d1f751dadbf2fd1594274bf1"
     },
     "config.json": {
       "bytes": 4336,
       "sha256": "405f0dcd92f7feba6011ed4eac5c8d4f74cba9712f07fd5cfa3063bbdd95402c"
     },
     "tokenizer.json": {
       "bytes": 3676772,
       "sha256": "f8984687e4a92a3503d521396d454b7d68e9fdaab2a0288eb3536c7c1aa4bc20"
     },
     "tokenizer_config.json": {
       "bytes": 5059,
       "sha256": "e78a3ebb48a0b9437efd1823b6b726c823da89e49dd8bcc90c02419d9baa772b"
     },
     "special_tokens_map.json": {
       "bytes": 1632,
       "sha256": "068594063e37662c02b21acf42ebb334ef6a74fb810e68a2368f88f08351de76"
     }
   }
   PYMANIFEST
   .venv/bin/python - <<'PYHASH'
   import hashlib,json,pathlib
   for name,expected in json.load(open('asset-manifest.json')).items():
    file=pathlib.Path('models')/name;digest=hashlib.sha256()
    with file.open('rb') as stream:
     for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
    assert file.stat().st_size==expected['bytes'] and digest.hexdigest()==expected['sha256'],name
   print('All six native asset length/hash checks passed.')
   PYHASH

   Expected result: All six native asset length/hash checks pass. The main model and AudioVAE are independently checked; no weight-body validation was performed by the author.

Configure

1. 1. Create a stable private local key

   Generate only when absent. Never print/share the key; preserve an existing key on repeated setup.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   test -s .api-key || { echo "Restore empty service key before continuing"; exit 1; }

   Expected result: A private nonempty key exists; the server requires exactly 64 lowercase hexadecimal characters.

2. 2. Create the bounded native speech harness

   This reader-authored serial loopback server accepts only its fixed sentence fixture ID. It loads local pinned assets offline, disables compilation/warm-up and adds a base-keyset check because the native loader uses strict=False for LoRA compatibility. No reference voice, denoiser, normalizer or voice clone is used. Native generator returns patch features for an exact early-stop cap check.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   cat > voxcpm2_tts.py <<'PYSERVER1'
   """Private fixed-sentence native VoxCPM2 research harness; no model imports in boundary tests."""
   import base64,hashlib,hmac,io,json,math,pathlib,struct,wave
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8994
   TEXT='The library opens at nine in the morning. Please return your books by Friday.'
   CAPACITY=8192
   MAX_TEXT=128
   MAX_PATCHES=256
   
   def validate_audio(result):
       samples,text_count,patches,cap=result
       if type(text_count) is not int or not 1<=text_count<=MAX_TEXT:raise ValueError('text_budget')
       if type(cap) is not int or cap!=min(text_count*6+10,MAX_PATCHES):raise ValueError('effective_cap')
       if type(patches) is not int or not 4<=patches<cap:raise ValueError('not_native_early_stop')
       if text_count+1+cap>CAPACITY:raise ValueError('shared_budget')
       if not isinstance(samples,list) or not 1<=len(samples)<=MAX_PATCHES*7680:raise ValueError('audio_length')
       if any(type(x) not in (float,int) or not math.isfinite(x) or abs(x)>1 for x in samples):raise ValueError('invalid_or_clipping_audio')
       rms=math.sqrt(sum(x*x for x in samples)/len(samples))
       if rms==0:raise ValueError('silent_audio')
       buf=io.BytesIO()
       with wave.open(buf,'wb') as out:
           out.setnchannels(1);out.setsampwidth(2);out.setframerate(48000)
           out.writeframes(b''.join(struct.pack('<h',round(x*32767)) for x in samples))
       return {'fixture':'author-sentence-v1','text':TEXT,'text_tokens':text_count,'audio_start_positions':1,
               'native_shared_sequence_capacity':CAPACITY,'acoustic_patches':patches,'effective_patch_cap':cap,
               'native_early_stop':True,'sample_rate':48000,'channels':1,'sample_count':len(samples),
               'duration_seconds':len(samples)/48000,'peak':max(abs(x) for x in samples),'rms':rms,
               'wav_base64':base64.b64encode(buf.getvalue()).decode('ascii'),'generated_text_tokens':None}
   
   def make_encoder():
       import os
       os.environ['HF_HUB_OFFLINE']='1';os.environ['TRANSFORMERS_OFFLINE']='1'
       import torch
       from safetensors import safe_open
       from voxcpm.model.voxcpm2 import VoxCPM2Model
       torch.set_num_threads(4)
       path=ROOT/'models'
       expected=json.loads((ROOT/'asset-manifest.json').read_text())
       for name,info in expected.items():
           file=path/name; digest=hashlib.sha256()
           with file.open('rb') as stream:
               for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
           if file.stat().st_size!=info['bytes'] or digest.hexdigest()!=info['sha256']:raise RuntimeError('asset_mismatch:'+name)
       model=VoxCPM2Model.from_local(str(path),optimize=False,device='cpu')
       # Native loader uses strict=False; this independent base-keyset guard is explicit.
       with safe_open(str(path/'model.safetensors'),framework='pt',device='cpu') as f:keys=set(f.keys())
       vae=torch.load(path/'audiovae.pth',map_location='cpu',weights_only=True)
       vae=vae.get('state_dict',vae);keys|={'audio_vae.'+k for k in vae}
       if keys!=set(model.state_dict()):raise RuntimeError('base_state_keyset_mismatch')
       del vae
       if model.device!='cpu' or model.config.dtype!='bfloat16' or model.config.max_length!=8192:raise RuntimeError('runtime_configuration')
       if model.sample_rate!=48000 or model.patch_size!=4 or model._decode_chunk_size!=1920:raise RuntimeError('audio_configuration')
       def encode():
           ids=model.text_tokenizer(TEXT);n=len(ids);cap=min(n*6+10,MAX_PATCHES)
           if not 1<=n<=MAX_TEXT or n+1+cap>CAPACITY:raise ValueError('input_budget')
           with torch.inference_mode():
               audio,target_ids,features=model.generate_with_prompt_cache(target_text=TEXT,prompt_cache=None,
                   min_len=2,max_len=MAX_PATCHES,inference_timesteps=10,cfg_value=2.0,
                   retry_badcase=False,retry_badcase_max_times=1,seed=1234)
           if target_ids.tolist()!=ids or list(features.shape[1:])!=[4,64]:raise RuntimeError('native_output_shape')
           if audio.ndim!=2 or audio.shape[0]!=1:raise RuntimeError('mono_shape')
           return audio[0].float().tolist(),n,int(features.shape[0]),cap
       return encode
   
   PYSERVER1

   Expected result: Dedicated wrapper is saved. Actual asset/keyset/model/dtype/shape checks run only when the reader starts the service; errors stop startup.

3. 3. Append the bounded HTTP boundary

   Append the serial authenticated fixed-input HTTP server and entry point to the same dedicated file.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   cat >> voxcpm2_tts.py <<'PYSERVER2'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(status)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               self.wfile.write(data)
               self.close_connection = True
   
           def authorized(self):
               values = self.headers.get_all('Authorization', [])
               return len(values) == 1 and hmac.compare_digest(values[0].encode(), ('Bearer ' + key).encode())
   
           def do_GET(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'native_shared_sequence_capacity': 8192,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/speech':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   return self.reply(413, {'error': 'body_limit'})
               if self.headers.get_content_type() != 'application/json':
                   return self.reply(415, {'error': 'json_required'})
               try:
                   body = self.rfile.read(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-sentence-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_audio(encoder())
               except Exception:
                   return self.reply(500, {'error': 'synthesis_or_audio_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__=='__main__':
       encoder=make_encoder()
       with create_server(encoder,(ROOT/'.api-key').read_text().strip()) as server:server.serve_forever()
   PYSERVER2

   Expected result: Complete native wrapper saved; no external model imports occur until reader start.

Start the model

1. 1. Start the private native TTS experiment

   Run in foreground with exact source path. The native local loader uses BF16 on CPU and FP32 AudioVAE, optimizeFalse, no reference or auxiliary denoiser. One request runs at a time; native static caches are refilled from the fixed text each time. No inference deadline is implemented.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   PYTHONPATH="$HOME/voxcpm2-tts/native-voxcpm/src" HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 .venv/bin/python voxcpm2_tts.py

   Expected result: After actual reader execution and successful asset/native checks, loopback 8994 listens. No successful load, audio or CPU fit is preclaimed.

Verify the setup

1. 1. Check the waveform and listen to the complete original sentence

   Use a second terminal on the server. Request only the fixed original sentence. Protocol checks validate native text/control/acoustic budgeting, strict early stop below its effective cap and finite non-silent unclipped PCM16 output. Listen to the generated WAV yourself: the sentence should be intelligible and complete, with no cut-off words or unwanted noise. Native early termination is not proof of spoken content or quality; record a failed listening result honestly.

   CommandCopy command

   set -e
   cd "$HOME/voxcpm2-tts"
   TTS_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8994/health -H "Authorization: Bearer $TTS_KEY"
   curl --fail http://127.0.0.1:8994/v1/speech -H "Authorization: Bearer $TTS_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-sentence-v1"}' -o speech.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8994/health
   unset TTS_KEY
   .venv/bin/python - <<'PYVERIFY'
   import base64,io,json,wave
   d=json.load(open("speech.json"));assert d["native_shared_sequence_capacity"]==8192 and d["generated_text_tokens"] is None
   assert d["native_early_stop"] is True and 4<=d["acoustic_patches"]<d["effective_patch_cap"]
   assert d["effective_patch_cap"]==min(d["text_tokens"]*6+10,256) and d["text_tokens"]+1+d["effective_patch_cap"]<=8192
   b=base64.b64decode(d["wav_base64"],validate=True)
   with wave.open(io.BytesIO(b),"rb") as w:
    assert w.getframerate()==48000 and w.getnchannels()==1 and w.getsampwidth()==2 and w.getnframes()==d["sample_count"]
   open("speech.wav","wb").write(b)
   print("Listen to speech.wav against:",d["text"])
   print("Native early stop is not a guarantee of complete intelligible speech.")
   PYVERIFY

   Expected result: Authenticated health/audio return documented JSON; missing key returns401. Waveform checks pass only after actual synthesis. Save speech.wav, listen and record whether the complete original sentence is intelligible. No exact voice, duration or quality result is preclaimed.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Source-documented setup only; no model weights, installation, build, model imports, speech inference, memory fit, latency or quality observations.
• Exact frozen API is VoxCPM2. Publisher describes 2B diffusion-autoregressive TTS with 30 languages, but this bounded original English sentence does not establish multilingual quality or family completeness.
• The model uses text tokenizer IDs and continuous acoustic features. 8,192 is configured shared base/residual cache capacity, not 8,192 text tokens; lm_config 32,768 positions is distinct. Actualtext+audio-start+effectivepatchcap are checked. Output texttokens are N/A.
• The native CPU path preserves BF16 model compute; AudioVAE is FP32. No claim of CPU BF16 hardware acceleration, FP32-equivalent behavior, fit or speed. Native internal scripted functions remain; optimizeFalse disables model compilation/warm-up.
• Both main safetensors and same-revision AudioVAE PyTorch checkpoint are pinned. Native AudioVAE loader uses weights_only=True. Native model load uses strict=False; the custom startup keyset guard separately rejects missing/unexpected base keys. This has not been executed with weights.
• Native zero-shot generate_with_prompt_cache(None) returns generated feature patches. With no reference/continuation context_len 0, fewer patches than min(textlength*6+10,256) proves the native early-break branch, not intelligibility or utterance correctness. Cap exhaustion is rejected.
• Ten diffusion steps, guidance 2, minimum length 2 and seed 1234 are explicit native-compatible settings. No text sampler or chat/thinking template applies. Stochastic reproducibility and speaker consistency are not benchmarked.
• Fixed-input HTTPServer has private auth, serial execution, request-body/socket limits and strict response checks. It has no model-execution deadline, public/TLS/multiuser suitability or voice safety certification. Privileged local users remain trusted.
• The direct dependency/source path avoids broad native UI/ASR/normalizer/denoiser integrations. Native code imports its originalVoxCPM compatibility module and torchaudio, so CPU torchaudio is still installed. Exact runtime compatibility is source-qualified, not installed/model-tested; pip freeze records resolved transitive dependencies.
• 32 GiB RAM, 40 GiB disk and four threads are planning allocations. Published weight sizes exclude caches/activations/audio buffers/build/OS. Dated AWS compute/storage source facts and unknown setup, egress, tax, local power and administration costs remain distinct.
• Reference voice/cloning requires separate consent/rights/input validation. Streaming, multilingual/voice-design/continuation, GPU and longer texts remain useful unqualified scope; this pair does not exhaust supported workflows.

Asset/keyset/model load error

Stop and verify exact model and AudioVAE assets, native source revision and dependency versions. Do not relax startup checks or substitute a different release/codec.

Audio cap reached, silence, clipping or failed listening

Record the task failure. Native early stopping, finite shape and successful HTTP do not prove intelligibility or sentence completion. Do not retry indefinitely or claim success from an invalid audio result.

Missing package, unsupported dtype or slow CPU

Save the exact traceback and resolved environment. BF16 CPU is native source support, not guaranteed fast execution or full compatibility. Changing dtype/device/dependencies requires requalification.

401 or rejected input

Use the exact fixed fixture JSON and matching private key, with loopback/SSH access. Do not expose the port or add arbitrary text/voice uploads without a new boundary assessment.

Stop the model

1. 1. Stop the native speech endpoint

   Press Ctrl+C in its foreground terminal and unset keys. Confirm loopback 8994 closes.

   CommandCopy command

   unset TTS_KEY

   Expected result: Python speech service stops; no active endpoint remains.

Uninstall

1. 1. Remove dedicated experiment files

   Back up desired outputs/licensing records, confirm the dedicated path and remove only this workspace. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/voxcpm2-tts"

   Expected result: Dedicated weights, source checkout, virtualenv, manifest, wrapper, audio and private key are removed.

Exact model and requirements

Artifact:

OpenBMB / VoxCPM2 native model with AudioVAE V2 auxiliary checkpoint · checkpoint openbmb/VoxCPM2

Publisher revision:

32279effe8c19989596f05d353d1447f51d9e915

Artifact digest:

sha256:f7f964cfa9da23653baec6e6f7750719977ad944ed9f95fe52fe3a620506891d

Download size:

4.27 GiB

Memory requirement (unknown):

Unknown — No full process peak measured for exact native BF16 CPU, FP32 AudioVAE, static caches and output buffers.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor or pipeline sharding. Whole native speech model, text-token preparation, acoustic patch generation and AudioVAE decoding run on CPU; no GPU or disk offload.

KV cache:

Native BF16 static base/residual LM caches, each capacity8192; AudioVAE/state and full process memory unmeasured.

Workload:

8,192 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native LlamaTokenizerFast plus mask_multichar_chinese_tokens; add audio-start control101 · 32279effe8c19989596f05d353d1447f51d9e915; runtime cce58a5b59303c9bd63d12f23afe6a49a4a80c59

Sampling / reasoning:

Native diffusion guidance cfg_value 2.0 and 10 inference steps, min_len 2, max_len 256 acoustic patches, seed 1234, retry_badcase False with one attempt. Native effective cap is min(text_token_count*6+10,256). No text temperature/top-p sampler applies. No chat template or reasoning output. contextTokens 8,192 labels native shared text/acoustic autoregressive positions. Reserve actual textcount+one audio-start+effective acoustic cap within 8,192; native lm_config 32,768 position setting is distinct from allocated cache. Generated acoustic feature count strictly below the effective cap establishes native early stop only for zero-shot context_len 0, not intelligibility/quality.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

The exact native model card explicitly releases VoxCPM2 under Apache-2.0; the native runtime includes the actual Apache-2.0 license text. Retain relevant license/copyright/NOTICE and modification notices; patent and trademark conditions apply. The AudioVAE is part of the same published model repository. This fixed original-text recipe supplies no reference voice or external audio and grants no rights to impersonate anyone. No legal clearance or publisher endorsement is claimed.

• audiovae.pth · 0.35 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• config.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• tokenizer.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• tokenizer_config.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915
• special_tokens_map.json · 0 GiB · revision 32279effe8c19989596f05d353d1447f51d9e915

• Main checkpoint 4,580,080,592 bytes plus AudioVAE 376,951,122 bytes = 4,957,031,714 published weight bytes, excluding all runtime/cache/activation/OS overhead.
• 32 GiB RAM, 40 GiB disk and four threads are author planning allocations, not measured minima or guaranteed fit. BF16 CPU support does not imply fast execution.

Install the runtime

1. 1. Create the exact native CPU research environment

   Install exact CPU wheels and pinned direct dependencies, then use the pinned native source checkout via an explicit PYTHONPATH without installing its broad UI/ASR package bundle. The archived runtime import closure supports this fixed zero-shot path. No UI, denoiser, text normalizer, VAD, prompt-audio loading or external model is enabled. Record transitive/system resolution and stop on pip check or imports/model-load errors; this stack is not executed by the author.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl libsndfile1
   mkdir -p "$HOME/voxcpm2-tts/models"
   cd "$HOME/voxcpm2-tts"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91' 'https://download-r2.pytorch.org/whl/cpu/torchaudio-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=dfb1ae1d7da1e869a6a6a315cc2b2652c43e3aabb5184da4d363d1b4bb2c86a4'
   .venv/bin/python -m pip install 'numpy==2.2.3' 'transformers==4.57.6' 'einops==0.8.1' 'pydantic==2.10.6' 'safetensors==0.5.3' 'huggingface-hub==0.36.0' 'librosa==0.11.0' 'soundfile==0.13.1' 'tqdm==4.67.1'
   git clone https://github.com/OpenBMB/VoxCPM.git native-voxcpm
   git -C native-voxcpm checkout --detach cce58a5b59303c9bd63d12f23afe6a49a4a80c59
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned checkout/hashed CPU wheels and direct dependencies resolve; pip check passes for the selected installed packages and resolution is saved. The native source is imported only at reader start, through explicit PYTHONPATH. No actual model load is preclaimed.

Download the model

1. 1. Download and hash all native model and AudioVAE assets

   Reader downloads only; the author archived metadata/source without weight bodies. Use exact native revision for both weights and tokenizer/config files. The manifest pins file bytes rather than mutable API counters. Stop on any failure; do not accept an alternate codec or revision.

   CommandCopy command

   set -e
   cd "$HOME/voxcpm2-tts"
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/model.safetensors' -o 'models/model.safetensors'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/audiovae.pth' -o 'models/audiovae.pth'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/config.json' -o 'models/config.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/tokenizer.json' -o 'models/tokenizer.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/tokenizer_config.json' -o 'models/tokenizer_config.json'
   curl --fail --location 'https://huggingface.co/openbmb/VoxCPM2/resolve/32279effe8c19989596f05d353d1447f51d9e915/special_tokens_map.json' -o 'models/special_tokens_map.json'
   cat > asset-manifest.json <<'PYMANIFEST'
   {
     "model.safetensors": {
       "bytes": 4580080592,
       "sha256": "f7f964cfa9da23653baec6e6f7750719977ad944ed9f95fe52fe3a620506891d"
     },
     "audiovae.pth": {
       "bytes": 376951122,
       "sha256": "94b5d51e107e0507d4acc976cfdadb64edd6fd06d1f751dadbf2fd1594274bf1"
     },
     "config.json": {
       "bytes": 4336,
       "sha256": "405f0dcd92f7feba6011ed4eac5c8d4f74cba9712f07fd5cfa3063bbdd95402c"
     },
     "tokenizer.json": {
       "bytes": 3676772,
       "sha256": "f8984687e4a92a3503d521396d454b7d68e9fdaab2a0288eb3536c7c1aa4bc20"
     },
     "tokenizer_config.json": {
       "bytes": 5059,
       "sha256": "e78a3ebb48a0b9437efd1823b6b726c823da89e49dd8bcc90c02419d9baa772b"
     },
     "special_tokens_map.json": {
       "bytes": 1632,
       "sha256": "068594063e37662c02b21acf42ebb334ef6a74fb810e68a2368f88f08351de76"
     }
   }
   PYMANIFEST
   .venv/bin/python - <<'PYHASH'
   import hashlib,json,pathlib
   for name,expected in json.load(open('asset-manifest.json')).items():
    file=pathlib.Path('models')/name;digest=hashlib.sha256()
    with file.open('rb') as stream:
     for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
    assert file.stat().st_size==expected['bytes'] and digest.hexdigest()==expected['sha256'],name
   print('All six native asset length/hash checks passed.')
   PYHASH

   Expected result: All six native asset length/hash checks pass. The main model and AudioVAE are independently checked; no weight-body validation was performed by the author.

Configure

1. 1. Create a stable private local key

   Generate only when absent. Never print/share the key; preserve an existing key on repeated setup.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   test -s .api-key || { echo "Restore empty service key before continuing"; exit 1; }

   Expected result: A private nonempty key exists; the server requires exactly 64 lowercase hexadecimal characters.

2. 2. Create the bounded native speech harness

   This reader-authored serial loopback server accepts only its fixed sentence fixture ID. It loads local pinned assets offline, disables compilation/warm-up and adds a base-keyset check because the native loader uses strict=False for LoRA compatibility. No reference voice, denoiser, normalizer or voice clone is used. Native generator returns patch features for an exact early-stop cap check.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   cat > voxcpm2_tts.py <<'PYSERVER1'
   """Private fixed-sentence native VoxCPM2 research harness; no model imports in boundary tests."""
   import base64,hashlib,hmac,io,json,math,pathlib,struct,wave
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8994
   TEXT='The library opens at nine in the morning. Please return your books by Friday.'
   CAPACITY=8192
   MAX_TEXT=128
   MAX_PATCHES=256
   
   def validate_audio(result):
       samples,text_count,patches,cap=result
       if type(text_count) is not int or not 1<=text_count<=MAX_TEXT:raise ValueError('text_budget')
       if type(cap) is not int or cap!=min(text_count*6+10,MAX_PATCHES):raise ValueError('effective_cap')
       if type(patches) is not int or not 4<=patches<cap:raise ValueError('not_native_early_stop')
       if text_count+1+cap>CAPACITY:raise ValueError('shared_budget')
       if not isinstance(samples,list) or not 1<=len(samples)<=MAX_PATCHES*7680:raise ValueError('audio_length')
       if any(type(x) not in (float,int) or not math.isfinite(x) or abs(x)>1 for x in samples):raise ValueError('invalid_or_clipping_audio')
       rms=math.sqrt(sum(x*x for x in samples)/len(samples))
       if rms==0:raise ValueError('silent_audio')
       buf=io.BytesIO()
       with wave.open(buf,'wb') as out:
           out.setnchannels(1);out.setsampwidth(2);out.setframerate(48000)
           out.writeframes(b''.join(struct.pack('<h',round(x*32767)) for x in samples))
       return {'fixture':'author-sentence-v1','text':TEXT,'text_tokens':text_count,'audio_start_positions':1,
               'native_shared_sequence_capacity':CAPACITY,'acoustic_patches':patches,'effective_patch_cap':cap,
               'native_early_stop':True,'sample_rate':48000,'channels':1,'sample_count':len(samples),
               'duration_seconds':len(samples)/48000,'peak':max(abs(x) for x in samples),'rms':rms,
               'wav_base64':base64.b64encode(buf.getvalue()).decode('ascii'),'generated_text_tokens':None}
   
   def make_encoder():
       import os
       os.environ['HF_HUB_OFFLINE']='1';os.environ['TRANSFORMERS_OFFLINE']='1'
       import torch
       from safetensors import safe_open
       from voxcpm.model.voxcpm2 import VoxCPM2Model
       torch.set_num_threads(4)
       path=ROOT/'models'
       expected=json.loads((ROOT/'asset-manifest.json').read_text())
       for name,info in expected.items():
           file=path/name; digest=hashlib.sha256()
           with file.open('rb') as stream:
               for chunk in iter(lambda:stream.read(1024*1024),b''):digest.update(chunk)
           if file.stat().st_size!=info['bytes'] or digest.hexdigest()!=info['sha256']:raise RuntimeError('asset_mismatch:'+name)
       model=VoxCPM2Model.from_local(str(path),optimize=False,device='cpu')
       # Native loader uses strict=False; this independent base-keyset guard is explicit.
       with safe_open(str(path/'model.safetensors'),framework='pt',device='cpu') as f:keys=set(f.keys())
       vae=torch.load(path/'audiovae.pth',map_location='cpu',weights_only=True)
       vae=vae.get('state_dict',vae);keys|={'audio_vae.'+k for k in vae}
       if keys!=set(model.state_dict()):raise RuntimeError('base_state_keyset_mismatch')
       del vae
       if model.device!='cpu' or model.config.dtype!='bfloat16' or model.config.max_length!=8192:raise RuntimeError('runtime_configuration')
       if model.sample_rate!=48000 or model.patch_size!=4 or model._decode_chunk_size!=1920:raise RuntimeError('audio_configuration')
       def encode():
           ids=model.text_tokenizer(TEXT);n=len(ids);cap=min(n*6+10,MAX_PATCHES)
           if not 1<=n<=MAX_TEXT or n+1+cap>CAPACITY:raise ValueError('input_budget')
           with torch.inference_mode():
               audio,target_ids,features=model.generate_with_prompt_cache(target_text=TEXT,prompt_cache=None,
                   min_len=2,max_len=MAX_PATCHES,inference_timesteps=10,cfg_value=2.0,
                   retry_badcase=False,retry_badcase_max_times=1,seed=1234)
           if target_ids.tolist()!=ids or list(features.shape[1:])!=[4,64]:raise RuntimeError('native_output_shape')
           if audio.ndim!=2 or audio.shape[0]!=1:raise RuntimeError('mono_shape')
           return audio[0].float().tolist(),n,int(features.shape[0]),cap
       return encode
   
   PYSERVER1

   Expected result: Dedicated wrapper is saved. Actual asset/keyset/model/dtype/shape checks run only when the reader starts the service; errors stop startup.

3. 3. Append the bounded HTTP boundary

   Append the serial authenticated fixed-input HTTP server and entry point to the same dedicated file.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   cat >> voxcpm2_tts.py <<'PYSERVER2'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(status)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               self.wfile.write(data)
               self.close_connection = True
   
           def authorized(self):
               values = self.headers.get_all('Authorization', [])
               return len(values) == 1 and hmac.compare_digest(values[0].encode(), ('Bearer ' + key).encode())
   
           def do_GET(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'native_shared_sequence_capacity': 8192,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/speech':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   return self.reply(413, {'error': 'body_limit'})
               if self.headers.get_content_type() != 'application/json':
                   return self.reply(415, {'error': 'json_required'})
               try:
                   body = self.rfile.read(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-sentence-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_audio(encoder())
               except Exception:
                   return self.reply(500, {'error': 'synthesis_or_audio_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__=='__main__':
       encoder=make_encoder()
       with create_server(encoder,(ROOT/'.api-key').read_text().strip()) as server:server.serve_forever()
   PYSERVER2

   Expected result: Complete native wrapper saved; no external model imports occur until reader start.

Start the model

1. 1. Start the private native TTS experiment

   Run in foreground with exact source path. The native local loader uses BF16 on CPU and FP32 AudioVAE, optimizeFalse, no reference or auxiliary denoiser. One request runs at a time; native static caches are refilled from the fixed text each time. No inference deadline is implemented.

   CommandCopy command

   cd "$HOME/voxcpm2-tts"
   PYTHONPATH="$HOME/voxcpm2-tts/native-voxcpm/src" HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 .venv/bin/python voxcpm2_tts.py

   Expected result: After actual reader execution and successful asset/native checks, loopback 8994 listens. No successful load, audio or CPU fit is preclaimed.

Verify the setup

1. 1. Check the waveform and listen to the complete original sentence

   Use a second terminal on the server. Request only the fixed original sentence. Protocol checks validate native text/control/acoustic budgeting, strict early stop below its effective cap and finite non-silent unclipped PCM16 output. Listen to the generated WAV yourself: the sentence should be intelligible and complete, with no cut-off words or unwanted noise. Native early termination is not proof of spoken content or quality; record a failed listening result honestly.

   CommandCopy command

   set -e
   cd "$HOME/voxcpm2-tts"
   TTS_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8994/health -H "Authorization: Bearer $TTS_KEY"
   curl --fail http://127.0.0.1:8994/v1/speech -H "Authorization: Bearer $TTS_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-sentence-v1"}' -o speech.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8994/health
   unset TTS_KEY
   .venv/bin/python - <<'PYVERIFY'
   import base64,io,json,wave
   d=json.load(open("speech.json"));assert d["native_shared_sequence_capacity"]==8192 and d["generated_text_tokens"] is None
   assert d["native_early_stop"] is True and 4<=d["acoustic_patches"]<d["effective_patch_cap"]
   assert d["effective_patch_cap"]==min(d["text_tokens"]*6+10,256) and d["text_tokens"]+1+d["effective_patch_cap"]<=8192
   b=base64.b64decode(d["wav_base64"],validate=True)
   with wave.open(io.BytesIO(b),"rb") as w:
    assert w.getframerate()==48000 and w.getnchannels()==1 and w.getsampwidth()==2 and w.getnframes()==d["sample_count"]
   open("speech.wav","wb").write(b)
   print("Listen to speech.wav against:",d["text"])
   print("Native early stop is not a guarantee of complete intelligible speech.")
   PYVERIFY

   Expected result: Authenticated health/audio return documented JSON; missing key returns401. Waveform checks pass only after actual synthesis. Save speech.wav, listen and record whether the complete original sentence is intelligible. No exact voice, duration or quality result is preclaimed.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Source-documented setup only; no model weights, installation, build, model imports, speech inference, memory fit, latency or quality observations.
• Exact frozen API is VoxCPM2. Publisher describes 2B diffusion-autoregressive TTS with 30 languages, but this bounded original English sentence does not establish multilingual quality or family completeness.
• The model uses text tokenizer IDs and continuous acoustic features. 8,192 is configured shared base/residual cache capacity, not 8,192 text tokens; lm_config 32,768 positions is distinct. Actualtext+audio-start+effectivepatchcap are checked. Output texttokens are N/A.
• The native CPU path preserves BF16 model compute; AudioVAE is FP32. No claim of CPU BF16 hardware acceleration, FP32-equivalent behavior, fit or speed. Native internal scripted functions remain; optimizeFalse disables model compilation/warm-up.
• Both main safetensors and same-revision AudioVAE PyTorch checkpoint are pinned. Native AudioVAE loader uses weights_only=True. Native model load uses strict=False; the custom startup keyset guard separately rejects missing/unexpected base keys. This has not been executed with weights.
• Native zero-shot generate_with_prompt_cache(None) returns generated feature patches. With no reference/continuation context_len 0, fewer patches than min(textlength*6+10,256) proves the native early-break branch, not intelligibility or utterance correctness. Cap exhaustion is rejected.
• Ten diffusion steps, guidance 2, minimum length 2 and seed 1234 are explicit native-compatible settings. No text sampler or chat/thinking template applies. Stochastic reproducibility and speaker consistency are not benchmarked.
• Fixed-input HTTPServer has private auth, serial execution, request-body/socket limits and strict response checks. It has no model-execution deadline, public/TLS/multiuser suitability or voice safety certification. Privileged local users remain trusted.
• The direct dependency/source path avoids broad native UI/ASR/normalizer/denoiser integrations. Native code imports its originalVoxCPM compatibility module and torchaudio, so CPU torchaudio is still installed. Exact runtime compatibility is source-qualified, not installed/model-tested; pip freeze records resolved transitive dependencies.
• 32 GiB RAM, 40 GiB disk and four threads are planning allocations. Published weight sizes exclude caches/activations/audio buffers/build/OS. Dated AWS compute/storage source facts and unknown setup, egress, tax, local power and administration costs remain distinct.
• Reference voice/cloning requires separate consent/rights/input validation. Streaming, multilingual/voice-design/continuation, GPU and longer texts remain useful unqualified scope; this pair does not exhaust supported workflows.

Asset/keyset/model load error

Stop and verify exact model and AudioVAE assets, native source revision and dependency versions. Do not relax startup checks or substitute a different release/codec.

Audio cap reached, silence, clipping or failed listening

Record the task failure. Native early stopping, finite shape and successful HTTP do not prove intelligibility or sentence completion. Do not retry indefinitely or claim success from an invalid audio result.

Missing package, unsupported dtype or slow CPU

Save the exact traceback and resolved environment. BF16 CPU is native source support, not guaranteed fast execution or full compatibility. Changing dtype/device/dependencies requires requalification.

401 or rejected input

Use the exact fixed fixture JSON and matching private key, with loopback/SSH access. Do not expose the port or add arbitrary text/voice uploads without a new boundary assessment.

Stop the model

1. 1. Stop the native speech endpoint

   Press Ctrl+C in its foreground terminal and unset keys. Close SSH tunnels; stopping Python does not stop VM/storage billing.

   CommandCopy command

   unset TTS_KEY

   Expected result: Python speech service stops; no active endpoint remains.

Uninstall

1. 1. Remove dedicated experiment files

   Back up desired outputs/licensing records, confirm the dedicated path and remove only this workspace. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/voxcpm2-tts"

   Expected result: Dedicated weights, source checkout, virtualenv, manifest, wrapper, audio and private key are removed.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when the Python service stops.